HEARTBEAT | BLENDER | DISTCC | SCTP | RAPIDMIND | APIs

LINUX JOURNAL

™
High-Performance
Network Programming
SCTP Multiple Associations
HIGH-PERFORMANCE COMPUTING

Since 1994: The Original Magazine of the Linux Community

Heartbeat | Blender | distcc | SCTP | RapidMind | RHEL Clusters | APIs

+
THE PRESENT
AND FUTURE
KING OF HPC
INTERVIEW WITH
RapidMind
HIGH-PERFORMANCE
COMPUTING
NOVEMBER 2007 ISSUE 163

U RHEL Cluster Suite

NOVEMBER 2007 | ISSUE 163
w w w. l i n u x j o u rn a l . c o m

U Heartbeat and High Availability $5.99US $6.99CAN

11
U High-Availability E-Mail with Active Directory
U Distribute Compiles with distcc
0 74470 03102 4
 SEE US AT ANY OF THESE SHOWS!

Infrastructure Mgt. World

Manage Any Data Center. Scottsdale, AZ - Booth TBD
Sept. 10 - 12
Anytime. Anywhere. VM World
San Francisco, CA - Booth 1113
Sept. 11 - 13

AFCOM Data Center World

Dallas, TX - Booth 436
Sept. 17 - 18

High Perf. on Wall Street

New York, NY - Booth 216
Sept. 17

IDC Enterprise Infra. Forum

New York, NY - Booth TBD
Sept. 20

Interface Salt Lake City

Salt Lake City, UT - Booth 309
Oct. 4

GEOINT
San Antonio, TX - Booth 374
Oct. 22 - 24

Interop New York Fall

New York, NY - Booth 543
Oct. 24 - 25

AFCEA Asia-PAC TechNet

Honolulu, HI - Booth 516
Nov. 4 - 9

Super Computing
Reno, NV - Booth 164
Nov. 12 - 15

LISA
Dallas, TX - Booth 200
Nov. 14 - 15

DaCEY Awards
Atlanta, GA
Nov. 15

Gartner Data Center Conf.

Las Vegas, NV - Booth TBD
Nov. 27 - 30

Interface Seattle
Seattle, WA - Booth 206
Nov. 28

Avocent builds hardware and software to access, manage and control any IT asset in
your data center, online or offline, keeping it, and your business, “always on”.

Visit us on our Remote Control Tour. For locations near

Avocent, the Avocent logo and The Power of Being There, are registered
you, go to www.avocent.com/remotecontrol. trademarks of Avocent Corporation. ©2007 Avocent Corporation.
CONTENTS NOVEMBER 2007
Issue 163

HIGH-PERFORMANCE
COMPUTING

IMAGE ©ISTOCKPHOTO.COM/NARVIKK

FEATURES
46 RED HAT ENTERPRISE LINUX CLUSTER SUITE 60 DISTRIBUTED COMPUTING WITH DISTCC
The trusty Red Hat cluster. Put your lazy machines to work.
Khurram Shiraz Jes Hall

52 GETTING STARTED WITH HEARTBEAT

Availability in a heartbeat. ON THE COVER
Daniel Bartholomew • High-Performance Network Programming, p. 68
• SCTP Multiple Associations, p. 74
56 BUILDING A SCALABLE HIGH-AVAILABILITY • Interview with RapidMind, p. 64
• RHEL Cluster Suite, p. 46
E-MAIL SYSTEM WITH ACTIVE DIRECTORY
• Heartbeat and High Availability, p. 52
AND MORE • High-Availability E-Mail with Active Directory, p. 56
Cyrus-IMAP to the rescue. • Distribute Compiles with distcc, p. 60
Jack Chongjie Xue

2 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Agility.
With Coyote Point, you'll never have to wonder if your
network is fast enough or flexible enough. You'll know it is.
From local to global load balancing, application
acceleration or ultimate network manageability, Coyote
Point leads the pack. We take the guesswork and difficulty
out of application traffic management. You won’t find
anything faster, smarter or more affordable.
Find out why more than 2,000 businesses rely on us to
maximize their infrastructure. Learn what Coyote Point
could mean to your Web and network traffic.
Write [email protected] or call 1-877-367-2696.

Coyotes respond with lightning speed.

Your network should, too.

Copyright © 2007 All rights reserved. www.coyotepoint.com

CONTENTS NOVEMBER 2007
Issue 163
COLUMNS INDEPTH
22 REUVEN M. LERNER’S 64 PICKING THE RAPIDMIND
AT THE FORGE How to get those cores pumping.
Thinking about APIs Nicholas Petreley

26 MARCEL GAGNÉ’S 68 HIGH-PERFORMANCE

COOKING WITH LINUX NETWORKING
Because Nothing Says High
PROGRAMMING IN C
Performance Like a Good Race
Make the most of your bandwidth.
Girish Venkatachalam

74 MULTIPLE ASSOCIATIONS
WITH STREAM CONTROL
TRANSMISSION PROTOCOL
Chat up SCTP.
Jan Newmarch

80 ROMAN’S LAW AND

FAST PROCESSING WITH
MULTIPLE CPU CORES 64 RAPIDMIND
32 DAVE TAYLOR’S Life in the -fast lane.
WORK THE SHELL Roman Shaposhnik
Keeping Score in Yahtzee

34 JON "MADDOG" HALL’S

BEACHHEAD
86 HIGH-PERFORMANCE
LINUX CLUSTERS
Next Month
Linux in the Top 500.
Navigating by the Sun David Morton LAPTOPS
38 DOC SEARLS’
88 OPEN-SOURCE COMPOSITING It has been a good year for
LINUX FOR SUITS innovation in Linux laptops
IN BLENDER
The Usefulness Paradigm with many new models,
Power compositing in Blender.
Dan Sawyer
technologies and brands now
96 NICHOLAS PETRELEY’S at your disposal. In next issue’s
/VAR/OPINION Laptop Buying Guide, we’ll tell
Is Hardware Catching Up to Java? you what to look for and where
to buy the Linux laptop that
best fulfills your needs and
IN EVERY ISSUE won’t bust your budget.
8 LETTERS
As always, there’s much more.
12 UPFRONT
Reuven M. Lerner will show
18 TECH TIPS you how to create your own
42 NEW PRODUCTS Facebook applications, and
81 ADVERTISERS INDEX Ben Martin will explain how
to protect your laptop files
with FUSE and an on-line
storage service.

USPS LINUX JOURNAL (ISSN 1075-3583) is published monthly by Belltown Media, Inc., 2211 Norfolk, Ste 514, Houston, TX
77098 USA. Periodicals postage paid at Houston, Texas and at additional mailing offices. Cover price is $5.99 US. Subscription rate is
$25/year in the United States, $32 in Canada and Mexico, $62 elsewhere. POSTMASTER: Please send address changes to Linux
Journal, PO Box 980985, Houston, TX 77098. Subscriptions start with the next issue.

4 | november 2007 w w w. l i n u x j o u r n a l . c o m
..;174%14'5#4'$'.10)6175T

&RUHVVWDUWLQJDW

T
'48'422.+#0%'5
LNNTMGGTKIEN
‹462/&20,1&$OOULJKWVUHVHUYHGDOOWUDGHPDUNVDUHWKHSURSHUW\RIWKHLUUHVSHFWLYHFRPSDQLHV
 Editor in Chief
Nick Petreley, [email protected]

Executive Editor Jill Franklin

[email protected]

GO SOLID.
Senior Editor Doc Searls
[email protected]
Art Director Garrick Antikajian

INCREASE RELIABILTY.
[email protected]
Products Editor James Gray
[email protected]
Editor Emeritus Don Marti
[email protected]
Technical Editor Michael Baxter
[email protected]
Senior Columnist Reuven Lerner
[email protected]
Chef Français Marcel Gagné
[email protected]
Security Editor Mick Bauer
[email protected]

Contributing Editors
David A. Bandel • Ibrahim Haddad • Robert Love • Zack Brown • Dave Phillips
Marco Fioretti • Ludovic Marcotte • Paul Barry • Paul McKenney • Dave Taylor

Proofreader Geri Gale

Publisher Carlie Fairchild

[email protected]

solid state systems General Manager Rebecca Cassity

[email protected]
fully x86 compatible Director of Sales Laura Whiteman
fanless, quiet operation Regional Sales Manager
[email protected]
Joseph Krack
[email protected]
Regional Sales Manager Kathleen Boyle
[email protected]

Circulation Director Mark Irgang

[email protected]
Marketing Coordinator Lana Newlander
[email protected]

System Administrator Mitch Frazier

[email protected]
Webmaster Katherine Druckman
[email protected]

Accountant Candy Beauchamp

[email protected]

Direct-Plug Linux Journal is published by, and is a registered trade name of, Belltown Media, Inc.
PO Box 980985, Houston, TX 77098 USA
IDE Flash Modules
Editorial Advisory Board
Daniel Frye, Director, IBM Linux Technology Center
Intel, VIA & AMD CPUs Jon “maddog” Hall, President, Linux International
Lawrence Lessig, Professor of Law, Stanford University
Ransom Love, Director of Strategic Relationships, Family and Church History Department,
Church of Jesus Christ of Latter-day Saints
(IÀFLHQF\5DWHG Sam Ockman
Bruce Perens
PicoPSU Power Supplies Bdale Garbee, Linux CTO, HP
Danese Cooper, Open Source Diva, Intel Corporation

Advertising
E-MAIL: [email protected]
URL: www.linuxjournal.com/advertising
DISCOVER MINI-ITX. PHONE: +1 713-344-1956 ext. 2

Subscriptions
E-MAIL: [email protected]
URL: www.linuxjournal.com/subscribe
PHONE: +1 713-589-3503
FAX: +1 713-589-2677
TOLL-FREE: 1-888-66-LINUX
MAIL: PO Box 980985, Houston, TX 77098 USA
Please allow 4–6 weeks for processing address changes and orders
PRINTED IN USA

LINUX is a registered trademark of Linus Torvalds.

www.logicsupply.com
 letters
years use more power and produce more waste forbids consumers to copy and give away
heat. Even a PIII might have cooling problems! music media. Their morality is based on laws
that were manufactured by a Congress eager
One thing that is not widely known is that a to be bribed. It’s easy to burst the RIAA’s
P4 will generally take more clock cycles than moral bubble. Just travel back to 1850 in
a PIII to complete a given amount of work— your mind and go hear a music performer.
a 1.5GHz P4 will often deliver the same
performance as a 1GHz PIII! If a machine In 19th-century America, the artists played and
is to run in a poorly ventilated space, a the people enjoyed. It had been that way for
PIII is a significantly better option than a P4. 10,000 years. If the artists were lucky or very
good, they could make a living by performing.
Hard drives take as much as 5W when Many do that to this very day. It is an honest liv-
idling. Use of other mass storage devices ing. They work at what they love. It’s their job.
(such as USB) might reduce the power use
for such a machine. It was a temporary quirk of technology that
allowed artists and record companies to turn
I have a small stockpile of old PIII machines for performances into commodities to be bought
occasions when I need a low-power machine. and sold. A few lucky artists could perform,
-- record and then sit back and let the bucks roll in
Your Wish Is Our Command Russell Coker while they snorted coke and bought mansions.
I would love to see a detailed explanation on That’s OK with me, and it still works for some.
the new Completely Fair Scheduler in the The Inevitable Descent into Anarchy But, the data is written on the wall, and it is an
Linux kernel, including areas where it may be The Linux OS is the most glamorous (I’m age-old warning that applies to all: adapt or die.
inferior to the existing scheduler and where it serious) example of [the] trend [toward data
excels. Also, I’d like to see an explanation of anarchy], but Linux is only an outgrowth of a Readers may wonder what all this has to do
the tickless patches, which apparently result much bigger universe of collaborative software with cell-phone networks. Carl Brown suggest-
in significant power savings. development that uses open-source licenses ed that there is nothing fundamentally block-
-- and techniques to leverage the talents and ing the concept of a user-created cell-phone
Chris Thompson creative impulses of hundreds of thousands network. It also is true that there is nothing
of people worldwide. Cheap MP3 players blocking users from building their own world-
We’ll look into a scheduling article, but we have and P2P file-sharing networks are another. wide data network, uncontrolled by govern-
an article on power savings in the works, includ- ments or corporations. If I can share my Wi-Fi
ing information on the tickless patches.—Ed. Direct evidence of the trend is plummeting CD with the neighbors in my apartment, I can
sales as users swap CDs, copy new ones, file bridge to the guy across the street who can
The Arcade Game Machine Is Hot! share and port their files to any device they run a directional shot to his buddy a quarter-
I was very interested to read Shawn Powers’ want. The efforts of the RIAA are becoming mile away who can...(ad infinitum).
article on building an arcade machine based laughable, as they sue a few unfortunates and
around Linux [LJ, August 2007]. However, one scold the rest of the population from a position The idea was not invented by me, although I
thing that concerns me is the lack of reference of pious hypocrisy. They are the biggest thieves humbly declare I thought of it myself. The idea
to cooling issues. If a PC is in a wooden box for of all. Also pathetic are the many pop stars, now is so obvious, anyone with an understanding of
any period of time without a fan to force cool grown rich and fat sucking on the corporate tit, networks is likely to conceive of it. A quick
air through, there would be a risk of heat dam- who parrot the RIAA’s line. It matters not. Google search brings to light that user-operated
age to the machine, particularly the CPU and and user-supported community networks are
the hard drive (www.coker.com.au/energy/ The only reason artists and record companies already providing free service to many people.
computer-power.html). ever had control over music media was that The ISPs don’t like it, but there’s nothing they
it was impractical for listeners to create their can do about it. Those simple networks, howev-
The above URL has the energy use of some own. Given that ability, most listeners will er, are only attachments to the Internet. They
computers I have owned at various times (the follow my own common-sense creed: when still depend on the corporate infrastructure.
idle tests were idle at a Linux prompt—idle at a I get the data, it’s my data, and I’ll do with
DOS prompt or the BIOS uses significantly more it what I damn well please. Another and related sign of the trend toward
power). You will note that a P4 machine uses anarchy is revealed in the project called
significantly more power than a PIII. I can only The RIAA hands down morality lessons while Netsukuku (netsukuku.freaknet.org). That
guess at the relative power requirements of a cracking the lawsuit whip in hope of putting project community has created a dæmon
PIII and the original game hardware, but I the file-sharing genie back in its bottle. They that uses network interfaces to communicate
expect that all PCs manufactured in the last ten pretend that there is some moral code that directly with the same dæmon in a connected

8 | november 2007 w w w. l i n u x j o u r n a l . c o m
 GO STRAIGHT TO THE SOURCE!
MORE PRODUCTS, BETTER SERVICE, GUARANTEED.

KEEP YOUR BUSINESS RUNNING SMOOTHLY

PROTECT YOUR SMALL BUSINESS WITH THE BUILT-IN SECURITY ENHANCEMENTS OF THE DUAL-CORE INTEL® XEON® PROCESSOR IN
YOUR SERVERSDIRECT SYSTEM.

SDR-1105T 1U ENTRY LEVEL SERVER EXCELLENT GENERAL PURPOSE SERVER FOR ORGANIZATIONS WITH THE NEED
FOR A LOW, ENTRY LEVEL PRICE
* 1U Rackmount Chassis with 520W power supply
* Supermicro X7DVL-L Server Board with Intel® 5000V (Blackford VS) Chipset
* Intel Quad-Core Xeon Processor E5310 1.6GHZ
* Total 512MB, 2pcs x 256MB Kingston DDR2 533Mhz FB-DIMM ECC
* Seagate SATAII 160GB 7200 RPM 8MB Cache SATA 3.0Gb/s Hard Drive

$1,129
* 4 x 1” Hot-swap SATA Drive Bays
* Two Intel® 82563EB Dual-port Gigabit Ethernet Controller
* Intel® ESB2 SATA 3.0Gbps Controller RAID 0, 1, 5, 10 support

SDR-2503T 2U APPLICATION SERVER

Highest performing with Dual Core/ Quad Core Xeon CPU based. Excellent with general purpose applications
and provide the most power.
* 2U Rackmount Chassis with 650W power supply * Seagate SATAII 250GB 7200 RPM 16MB Cache SATA 3.0Gb/s Hard
* Supermicro X7DVL-E Server Board with Intel® 5000V (Blackford VS) Drive
Chipset * 6 x 1” Hot-swap SATA Drive Bays

$1,259
* Intel Quad-Core Xeon Processor E5310 1.6GHZ * Intel® (ESB2/Gilgal) 82563EB Dual-port Gigabit Ethernet Controller
STARTING
* Total 512MB, 2pcs x 256MB Kingston DDR2 533Mhz FB-DIMM ECC * Intel® ESB2 SATA 3.0Gbps Controller RAID 0, 1, 5, 10 support
PRICE
----------------------------------------------------------------------- ----------------------------------
SDR-3500T 3U DATABASE SERVER
Easily Scalable storage solution with hot-swap functionality for growing businesses
* 3U Rackmount chassis with Redundant 800W power supply * Seagate SATAII 500GB 7200 RPM 16MB Cache SATA 3.0Gb/s
* Supermicro X7DBE+ Server Board with Intel® 5000P (Blackford) Hard
Chipset Drive
* Intel Quad-Core Xeon Processor E5310 1.6GHZ * 16 x 1" Hot-swap SATA Drive Bays
* Total 1024MB, 2pcs x 512MB Kingston DDR2 533MHz FB-DIMM * Dual-port Gigabit Ethernet Controller
STARTING
PRICE $1,999 ECC

----------------------------------------------------------------------- ----------------------------------
* Intel ESB2 SATA 3.0Gbps Controller RAID 0, 1, 10 support

SDR-7045B-TB 4U FILE SERVER

4U Quad-Core Xeon Server offers excellent value and expandability
* 4U Rackmountable / Tower with 650W power supply * Seagate SATAII 750GB 7200 RPM 16MB Cache SATA 3.0Gb/s
* Supermicro Super X7DBE Server Board with Intel® 5000P Hard Drive
(Blackford) Chipset * 6 x 3.5" Hot-swap SAS/SATA Drive Bays
* Intel Quad-Core Xeon Processor E5310 1.6GHZ * Dual-port Gigabit Ethernet Controller
* Total 1024MB, 2pcs x 512MB Kingston DDR2 667MHz FB-DIMM * Intel ESB2 SATA 3.0Gbps Controller RAID 0, 1, 5, 10 support
STARTING
PRICE $1,199 ECC

----------------------------------------------------------------------- ----------------------------------
SDR-5111T 5U ADVANCED STORAGE SERVER
Quad Core dual Xeon CPU based, with 24 hot-swap hard disk bays suitable for 18TB of pure data Storage
capacity
* 5U Rackmount chassis with Redundant 1350W power supply * Seagate 750GB 7200 RPM 16MB Cache SATA 3.0Gb/s Hard Drive
* Supermicro X7DBE Server Board with Intel® 5000P (Blackford) * 24 x 1" Hot-swap Drive Bays
Chipset * Intel® (ESB2/Gilgal) 82563EB Dual-
* Intel Quad-Core Xeon Processor E5310 1.6GHZ port Gigabit Ethernet Controller
STARTING
$3,099
* Total 1024MB, 2pcs x 512MB Kingston DDR2 667MHz FB- * Intel ESB2 SATA 3.0Gbps Controller
PRICE DIMM ECC RAID 0, 1, 5, 10 support

SERVERS DIRECT CAN HELP YOU CONFIGURE YOUR NEXT HIGH PERFORMANCE SERVER SYSTEM - CALL US TODAY!
Our flexible on-line products configurator allows you to source a custom solution, or call and our product experts are standing by to
help you assemble systems that require a little extra. Servers Direct - your direct source for scalable, cost effective server solutions.

1.877.727.7887 | www.ServersDirect.com
Intel, Intel logo, Intel Inside, Intel Inside logo, Intel Centrino, Intel Centrino logo, Celeron, Intel Xeon, Intel SpeedStep, Itanium, Pentium, and Pentium III Xeon are trademarks of
Intel Corporation or it’s subsidiaries in the United States and other countries.
[ LETTERS ]

computer. In theory, the mesh network so The Ultimate Linux Box had nothing to
created can mimic the Internet’s TCP/IP do with Windows games or serving up
layer with a decentralized domain name DirectX 10, which isn’t even very useful on
system that uses distributed architecture Windows yet. We’ll consider your desires
and builds routes that connect the for the next ULB issue though.—Ed.
computers attached to its network. It
is, as the creators say, anarchical. Protection Money
At Your Service
I have to congratulate you on your
I do not claim to see the future. I only /var/opinion “The Benevolent Racketeer”
extrapolate what seems to be an inex- [LJ, August 2007]. MAGAZINE
orable transfer of communications and
PRINT SUBSCRIPTIONS: Renewing your
data management power from the elite to Your imaginary letter writer speaks right subscription, changing your address, paying your
the masses. If you believe in people, this out of the soul of everybody. The crowd invoice, viewing your account details or other
can only be a good thing. If you fear peo- reading the lines will equally agree and subscription inquiries can instantly be done on-line,
ple, or begrudge them control over their see their points addressed as the crowd www.linuxjournal.com/subs. Alternatively,
within the U.S. and Canada, you may call
own lives, then you will fight it...and lose. reading between the lines. Luckily, I
us toll-free 1-888-66-LINUX (54689), or
-- paid my insurance money to Linux internationally +1-713-589-2677. E-mail us at
Tim Copeland Journal, so you will not sue me for [email protected] or reach us via postal mail,
reading this terrific piece of word code. Linux Journal, PO Box 980985, Houston, TX
Practical Linux Box 77098-0985 USA. Please remember to include your
-- complete name and address when contacting us.
I eagerly awaited the current issue of LJ Kurt
[September 2007] to read the Ultimate DIGITAL SUBSCRIPTIONS: Digital subscriptions
Linux Box article. I plan to build a system mkdir Errors of Linux Journal are now available and delivered as
PDFs anywhere in the world for one low cost.
very soon and hoped to get some ideas. I too have a small nit to pick with
Visit www.linuxjournal.com/digital for more
respect to sample code and the lack of information or use the contact information above
Unfortunately, I just skimmed the article and error handling. It is trivial to add at least for any digital magazine customer service inquiries.
was very disappointed. What I really would some minimal error handling. In the
LETTERS TO THE EDITOR: We welcome
like to see is a Practical Linux Box. No, your September 2007 issue’s Letters, Jack your letters and encourage you to submit them
sidebar “Penultimate” doesn’t address my points out that mkdir error handling is to [email protected] or mail them to
needs any more than the ULB. Your ULB is non-existent. Bash scripts are notorious Linux Journal, 1752 NW Market Street, #200,
really the Ultimate [Windows] Gaming Box, for not doing any error handling. As Seattle, WA 98107 USA. Letters may be edited
for space and clarity.
isn’t it? For example, look at the Display Card for the mkdir issue that Jack points out,
section. It’s the latest and greatest DirectX 10 I would suggest the following: WRITING FOR US: We always are looking
card? How is this a good Linux box? for contributed articles, tutorials and real-
mkdir -p $dimension world stories for the magazine. An author’s
guide, a list of topics and due dates can be
The things I’m looking for may not be if [ $? -ne 0 ]; then
found on-line, www.linuxjournal.com/author.
what everyone else is looking for, but I’d echo "Error: could not create\
like to think there are enough people to directory $dimension" ADVERTISING: Linux Journal is a great
resource for readers and advertisers alike.
warrant some practical thinking in this return 1
Request a media kit, view our current
type of article, such as: fi editorial calendar and advertising due
dates, or learn more about other advertising
I Quiet/fanless power supply and case. The -p switch to mkdir will create all and marketing opportunities by visiting us
directories in the $dimension string on-line, www.linuxjournal.com/advertising.
Contact us directly for further information,
I Small case with enough room for DVD recursively and will not complain if any [email protected] or +1 713-344-1956 ext. 2.
and two hard drives. or all directories already exist. The main
reason for failure will be user permis-
I Affordable: $2,000–$4,000 is not prac- sions. This is handled by echoing the
tical or affordable. error and returning an error value to ON-LINE
the caller. WEB SITE: Read exclusive on-line-only content on
I Onboard video is ok. Linux Journal’s Web site, www.linuxjournal.com.
If the goal of these sample scripts is to Also, select articles from the print magazine
You could argue that I want a media PC, instruct beginners to learn techniques are available on-line. Magazine subscribers,
digital or print, receive full access to issue
but that’s not really my goal. I’d settle for with bash, error handling cannot be archives; please contact Customer Service for
one. I’m estimating that I can build a sys- ignored. The same goes for Ruby, further information, [email protected].
tem that fits the above specs for about Python, Perl, C and C++. The code
FREE e-NEWSLETTERS: Each week, Linux
$700, without monitor. Other names for samples in the SCTP (why does that
Journal editors will tell you what's hot in the world
what I want might be the Quiet Linux Box make me think of John Candy) article of Linux. Receive late-breaking news, technical tips
or the Affordable Linux Box. are excellent in this regard. and tricks, and links to in-depth stories featured
on www.linuxjournal.com. Subscribe for free
-- --
today, www.linuxjournal.com/enewsletters.
JT Moree steeve

10 | november 2007 w w w. l i n u x j o u r n a l . c o m
 UPFRONT NEWS + FUN
LJ Index,
November 2007
1. Average measured speed in MBps of a
broadband connection with “up to 8Mbps”
download speed: 2.7
Linus Torvalds mailing list tend to assume—for
diff -u has expressed keen
interest in finding
example, how to do a binary search
with git to identify precisely when
2. Lowest measured speed in KBps of a
broadband connection with “up to 8Mbps”
WHAT’S NEW someone to put a particular bug was introduced into download speed: 90
IN KERNEL together a full git the tree.
3. Number of consumers out of five who get
DEVELOPMENT repository of the Several projects have changed
the broadband speed they signed up for: 1
kernel, going all hands recently. Valerie Henson has
the way back to version 0.01. He’s had to abandon the Tulip driver, and 4. Percent of surveyed consumers who have
tried this himself a couple times, and now it looks like Kyle McMartin may felt misled by providers’ advertising: 30
other folks have made various efforts, become the official maintainer. Wim
but it’s a hard problem. Certainly, it Van Sebroeck has submitted a patch 5. Billions of Internet users in 2006: 1.1
would not be possible to include every to make Mike Frysinger the official
6. Additional millions of Internet users expected
patch that went into the kernel, in the maintainer of the Blackfin Watchdog
by 2010: 500
order it was included, because many driver. Mike Sharkey of Pike
patches never were sent to any public Aerospace Research Corporation 7. Millions of video streams per day served by
forum. Even finding the release has volunteered to take over the YouTube: 100
announcements for all the numbered otherwise unmaintained Parallel Port
versions will be difficult, and some driver on behalf of his company. 8. Number of surveillance cameras in London: 200
official releases are thought to be lost And, Anton Vorontsov recently
9. Trillions of bits sent by London surveillance
as well. It’s a daunting task, but a very became a co-maintainer of the cameras to their data center: 64
valuable one, even if done incompletely. Power Supply subsystem, along
If someone can do it, Linus has offered with David Woodhouse. 10. Terabytes accumulated per day by Chevron: 2
to comment the various early patches Over time, various features have
and releases, from memory. gone into the kernel to support more 11. Total exabytes of data in 2006: 161
Mingming Cao has submitted and more modern architectures. But, for
12. Multiple in millions of 2006 data total to
patches to allow the ext4 filesystem some of these features that have no seri- all information in all books ever written: 3
to perform checksum calculations on ous negative impact on older hardware,
its journal to make sure that any cor- such as the 386 processor, there’s been 13. Percentage of the digital universe that will
ruption is identified as quickly as possi- no real effort to isolate the unneeded be created by individuals by 2010: 70
ble. With interest from various folks, features from that older hardware.
including Andrew Morton, it looks Kernels compiled for those systems, 14. Percentage of the current digital universe
that is subject to compliance rules and
like this feature quickly will be adopted therefore, have tended to have larger
standards: 20
into the official tree, although ext4 still and larger binaries and to require more
remains a fairly experimental filesystem. and more RAM to run. For the most 15. Percentage of the current digital universe
LinuxConf Europe 2007 (LCE) will part, no one notices or cares, because that is potentially subject to security
host a semi-formal discussion of con- most people don’t bother running Linux applications: 30
tainers within the kernel. Serge E. on the 386 anymore. But, the effect has
16. Exabytes of “user-generated content”
Hallyn recently announced plans to been there, building gradually.
expected by 2010: 692
arrange for a conference room (includ- Jonathan Campbell recently
ing phone lines for anyone who can’t started submitting patches to ensure 17. Total exabytes of data expected by 2010: 988
be present but still wants to partici- that architectures like the 386 would
pate) and a series of half-hour presen- compile only features that actually 18. Percentage of the 2010 digital universe for
tations. Containers provide a way to would work on those systems. So, which organizations will be responsible for
security, privacy, reliability and compliance: 85
cluster processes into specific name- things like the Pentium TSC register
spaces that are isolated from the rest would not be included in compiled 19. Exabyte capacity of media ready to store
of the system and are related to 386 kernel binaries. The result of his newly created and replicated data in the
virtualization projects like Xen. work was a much smaller binary, and 2010 digital universe: 601
Michal Piotrowski has announced his patches probably will be adopted
the “Linux Kernel Tester’s Guide”, into the main kernel tree. This kind of 20. Year in which the amount of information
translated by Rafael J. Wysocki, at support for legacy systems might have created will surpass available storage
capacity for the first time: 2007
www.stardust.webpages.pl/files/ an impact on projects to bring com-
handbook/handbook-en-0.3-rc1.pdf. puting resources to third-world coun- Sources: 1, 2: which.co.uk, sourced by David
It is a long document representing tries and impoverished neighborhoods Meyer for ZDNet UK | 3, 4: moneysupermarket.com,
much work, it reads like a book, and or lower the cost of experimenting sourced by David Meyer for ZDNet UK | 5–20:
it clearly explains a lot of material that with clustered solutions. “Expanding the Digital Universe”, by John F.
most discussions on the linux-kernel —ZACK BROWN Gantz, et al., a March 2007 IDC whitepaper

12 | november 2007 w w w. l i n u x j o u r n a l . c o m
[ UPFRONT ]

From ACCESS to Excess

The first shoe dropped in September 2005, Application framework that fills in some of
when ACCESS Co. Ltd. of Japan announced the formerly missing APIs for mobile applica-
that it would acquire Palm OS developer tions. Hiker originally was available through
PalmSource for $324 million. The next shoe the MPL (Mozilla Public License), but it
dropped in February 2006, when PalmSource reportedly will be dual-licensed with the
detailed the ACCESS Linux Platform (ALP), as LGPL (v2) license as well.
an environment open to running Palm OS Looking beyond the superficial resem-
binaries and Java applications, in addition to blances between what ACCESS showed and
native Linux apps. Enough other shoes have the now-familiar iPhone, it’s clear that the
dropped since then to give the clear message mobile application market will divide between
that ALP has legs. The latest was at Web-based (iPhone) and native OS-based
LinuxWorld Expo in August 2007, when the (ACCESS, OpenMoko, maemo)—with the
company showed screenshots of an iPhone- latter embracing legacy apps developed for
like UI and provided more details around its other platforms as well. Thus, the pavement
plan to make Linux the most supportive envi- gets wider on the road to truly open mobile
ronment for mobile device and application devices and markets that grow on them.
development. Chief among these is the Hiker —DOC SEARLS

Raising the Social Tide

Brad Fitzpatrick has a long pedigree for a young entry (same as its “social graph” entry) said,
guy. Besides creating LiveJournal—one of the “Social network analysis views social relation-
earliest and most popular blogging services (and ships in terms of nodes and ties. Nodes are
an open-source one at that)—he is the creator the individual actors within the networks, and
of OpenID, Perlbal, MogileFS, memcached, ties are the relationships between the actors.”
djabberd and many other fine hacks. Astute Elsewhere in Wikipedia, “Graph” is explained
readers may recall “Distributed Caching with as a mathematical concept, “a set of objects
Memcached”, which Brad wrote for Linux called points, nodes, or vertices connected by
Journal in 2004. Today, memcached is one of links called lines or edges”.
the world’s most widely used distributed caching So, the first idea is to move the center of
methods, while OpenID is a breakaway leader in social networking gravity outside the silo’d sites,
the user-centric identity field. each of which impose inconveniences on them-
In August 2007, Brad published “Thoughts selves as well as their members. As Brad says,
on the Social Graph” (bradfitz.com/ “People are getting sick of registering and re-
social-graph-problem), which presents declaring their friends on every site.” The pro-
this problem statement: cess is also silly. Not long after Google’s Orkut
social network site went up, Rael Dornfest
There are an increasing number of new made fun of the friendship-declaring protocol
“social applications” as well as tradi- by walking up to people he knew and saying,
tional applications that either require “You are my friend? Yes or no?”
the “social graph” or that could pro- The second idea is to make society itself
vide better value to users by utilizing a platform. Or, in Brad’s terms, to “make the
information in the social graph. What social graph a community asset”, and “to build
I mean by “social graph” is the the guts that allow a thousand new social
global mapping of everybody and applications to bloom”.
how they’re related, as Wikipedia Significantly, most social network sites (all
describes it (en.wikipedia.org/wiki/ but MySpace, I believe) run on Linux. Wikipedia
Social_graph). Unfortunately, there too. If Brad’s right, we can equip the rising social
doesn’t exist a single social graph (or network tide that lifts all boats. There should
even multiple graphs that interoperate) even be plenty of work converting today’s silos
that’s comprehensive and decentralized. into tomorrow’s arks.
Rather, there exists hundreds of disperse For more, visit bradfitz.com/
social graphs, most of dubious quality social-graph-problem, or just look up “social
and many of them walled gardens. graph” on the vast Linux hack called Google.
The lucky top result probably will be Brad’s.
At the time, Wikipedia’s “social network” —DOC SEARLS
 [ UPFRONT ]

Bring SMS to the Live

Web with a FoxBox
storing messages (it comes default with
a 512MB card). It runs on the 2.6 Linux
kernel, has a BIS module for failover to
up to two backup appliances and an
internal clock with a backup battery
and NTP support.
By the way, you might remember
Acme as the source not only of the tiny
Acme Fox embedded Linux system
board, but for its optional Tux Case as
Acme’s SMS FoxBox well. Today, that also can come with an
Acme Fox SBC (single-board computer).
The forces at Acme Systems and KDev, With Kdev’s FoxServe firmware, it can
two veteran embedded Linux system work as a dynamic Web server (Apache,
and software developers based in Italy, PHP, SQLite, SSL/TLS and so on).
have conspired to bring the world SMS
FoxBox. It’s a Linux-based box dedicated
to sending and receiving SMS messages
that can be managed through a Web
interface. It can handle up to 30 incom-
ing messages at a time on a common
SIM card. It also works as an SMS to
TCP/IP gateway, so you can combine
SMS messaging with network and user
applications. You can SMS to and from
e-mail, MySQL, Web scripts, desktop
widgets, whatever.
SMS FoxBox is ideal for use in Live
Web conditions. Broadcasters can use
it to interact with their audiences. Acme’s Tux Case (Showing Separate Parts)
Emergency services can use it to flow
live reports onto Web sites or broad-
casts. Lightweight monitoring, alarm Resources
sending, trouble ticketing and remote Acme Systems: www.acmesystems.it
device management can be moved to KDev: www.kdev.it/joomla
SMS from other methods. Databases Company Visuals:
and address books can be kept current. www.kdev.it/HTDOCS/press.html
The unit comes with a GSM quad-
band modem, an SD/MMC card for —DOC SEARLS
[ UPFRONT ]

Frisky N800 Hack

Keijo Lähetkangas has a fun job at Nokia: what you like. Also, Aibo had no eyes. Puppy
taking the company’s N800 handheld com- not only has eyes on his display, but he also
puter (subject of our September 2007 cover can look at the world through the N800’s
story) and turning it into the brains, faces and built-in camera, which is in a barely noticeable
controllers of Robot pets. His Puppy and a black cylinder that pops out the side of the
four-wheeled Rover companion were stars in N800. It got confusing to take pictures of
the demo room at this year’s Guadec confer- Puppy and Rover while they took pictures
ence in Birmingham (UK). Puppy is remark- of me, which appeared on other N800s
ably flexible and expressive. It walks, sits, over wireless connections.
smiles, dances, sleeps and even—how can we Puppy and Rover’s non-N800 hardware
put this politely?—lifts one leg (meaning that is all from the Robotics.com catalog. The
Puppy is a male, or at least acts like one). controlling software, all open source, is at
Several Guadec attendees compared https://garage.maemo.org/projects/robot.
Puppy with Sony’s Aibo, the most notable You also can see Puppy in action at
difference (besides looks) being the Puppy’s youtube.com/puppyrobot.
openness. It’s a pure DIY pet—make of him —DOC SEARLS

They Said It
If our customers buy bandwidth from us, and they want to share it with
neighbors, or publicly, that doesn't make them bad customers.
—Joe Plotkin, Bway.net, www.businessweek.com/technology/content/jul2002/
tc2002073_1130.htm

...the commercial concerns from the very beginning, even when they were small,
were really very important. The commercial distributions were what drove a lot
of the nice installers, and pushed people to improve usability....I think commer-
cial users of Linux have been very important in actually improving the product.
I think all the technical people who have been involved have been hugely
important, but I think that the kind of commercial use that you can get with the
GPLv2 is also important—you need a balance between pure technology, and the
kinds of pressures you get from users through the market....If you have a purely
marketing (or customer) driven approach, you end up with crap technology in
the end. But I think that something that is purely driven by technical people will
also end up as crap technology in the end, and you really need a balance here.
—Linus Torvalds, www.linuxworld.com/news/2007/
080907-torvalds-on-linux-ms-softwares.html?page=2

It's exciting to go to work each day knowing that scads of companies are using
your software, then contacting you to get additional value. It's not easy by any
stretch, but it's a lot more efficient and productive than the proprietary model.
—Matt Asay, blogs.cnet.com/
8301-13505_1-9762210-16.html?part=rss&tag=feed&subj=TheOpenRoad
 "SF
ZPV
4IPDLFE
CZ
UIF

IJHI
DPTU

PG
J4$4*

'JCSF
$IBOOFM

4"/
TUPSBHF

"P&
JT
UIF
BOTXFS
"5"PWFS&UIFSOFU

'BTU
3FMJBCMF
4JNQMF
TUPSBHF
XXXDPSBJEDPN


'BTU

(JHBCJU
&UIFSOFU
4UPSBHF

XJUIPVU
UIF

5$1*1
PWFSIFBE


6OMJNJUFE
FYQBOEBCJMJUZ
BU
UIF


MPXFTU
QPTTJCMF
QSJDF
QPJOU

¥ 
:PV
XBOU
NPSF
TUPSBHFyZPV




&UIFS%SJWF
43YYYY
KVTU
CVZ
NPSF
EJTLT
o
JUT
UIBU


r
'BTU

'MFYJCMF
3"*%
BQQMJBODFT

TJNQMF
XJUI
TMPUT
GPS
IPU
TXBQ
4"5"
EJTLT
¥


r
$IFDL
PVU
PVS
GVMM
MJOF
PG
&UIFS%SJWF 4UPSBHF
BOE



7JTJU
VT
BU
XXXDPSBJEDPN
7JSUVBM4UPSBHF
"QQMJBODFT
BOE
/"4
(BUFXBZT

 XXXDPSBJEDPN
 TECH TIPS
Display date and time in the past or future and pop through directories easily.

UU Show Date or Time, Past or Future I -f format: Display the date and time according to formatting
When working on Linux/UNIX platforms, I frequently find it useful directives specified in format.
to obtain the date and time in the past or future. Whether
scheduling jobs, searching for files from a certain date or deter- Options e and f are incompatible. Specifying them together
mining the day on which a certain date falls, countless scenarios on the command line terminates the program abnormally. The
need a routine to compute and display the date and time in the default output of showdate can be tweaked with the formatting
past or future. I searched for a suitable program, but ultimately directives and argument to -f, which are identical to the ones
had to write one myself. This program is called showdate. It is used by the standard date command. The important thing is that
written in the C language originally on UNIX and has been ported all computations are performed by taking either a positive (future)
over to Linux as well. You can download the code from the LJ FTP or negative (past) offset from the current date and time (now),
site: ftp.linuxjournal.com/pub/lj/listings/issue163/9877.tgz. which is its datum.
After obtaining the source code, assemble the showdate A good way to become familiar with any tool quickly is to under-
executable using an ANSI C compiler (either cc or gcc) as shown: stand how it is used. For example, the command to display the date
and time ten years ago, relative to now, would be (output of showdate
# cc showdate.c -o showdate has been omitted as the results depend on the value of now):

Store the executable in /usr/local/bin or a directory of your choice, # showdate -y -10

and invoke it without any arguments to obtain usage:
To find out the epoch seconds elapsed for the above scenario, use:
# showdate
usage: showdate # showdate -y -10 -e
[-y [+|-]years]
[-m [+|-]months] A futuristic date of five years, two months and 23 days from
[-d [+|-]days] now in the YY-MM-DD format would be output as shown below.
[-h [+|-]hours] The plus sign is optional for future dates and the two forms of the
[-M [+|-]minutes] command line below are equivalent (the minus sign is mandatory
[-s [+|-]seconds] for past dates):
[-e | -f format]
# showdate -y +5 -m +2 -d +23 -f %Y-%m-%d
showdate recognizes the following command-line options
and arguments: # showdate -y 5 -m 2 -d 23 -f %Y-%m-%d

I -y [+|-]years: Number of years in the past (-) or future (+) offset The options can appear in any order, as long as their contextual
from the current year. usage is unambiguous; therefore, the command line above could be
written as:
I -m [+|-]months: Number of months in the past (-) or future (+)
offset from the current month. # showdate -m 2 -f %Y-%m-%d -d 23 -y 5

I -d [+|-]days: Number of days in the past (-) or future (+) offset from The +- offsets can be combined in a single command line; howev-
the current day. er, mixing them up can lead to unexpected and erroneous results. If
now is January 1st 2003 12:00:00 AM UTC, showdate outputs:
I -h [+|-]hours: Number of hours in the past (-) or future (+) offset
from the current hour. # showdate -m -8 -M 32
Wed May 1 00:32:00 2002
I -M [+|-]minutes: Number of minutes in the past (-) or future (+)
offset from the current minute. The above command displays the date and time in the past—
eight months ago but 32 minutes from now, while the one below
I -s [+|-]seconds: Number of seconds in the past (-) or future (+) displays the date and time in the future—8 months from now but
offset from the current second. 32 minutes ago:

I -e: Display the time elapsed in seconds since the UNIX epoch # showdate -m 8 -m -32
(January 1, 1970 UTC). Sun Aug 31 23:28:00 2003

18 | november 2007 w w w. l i n u x j o u r n a l . c o m
 TECH TIPS

The capabilities of showdate can be leveraged to specify sub- $ dirs -l -p -d

minute job scheduling times. A quick way to schedule a batch job 12
minutes and 35 seconds from now would be: Suppose you make all the required changes in the first directory
(/home/sangeeth/soft/release2/src/). To remove that directory entry
# showdate -M 12 -s 35 -f %Y%m%d%H%M.%S | xargs at -f job-file -t from the list of directories that you created earlier, do the following:

The current date and time is tracked as the number of seconds $ popd
that have elapsed since the epoch. This number is stored in a signed
long, which means that on a 32-bit system, the timekeeping will break The above command removes the topmost directory entry
on Tuesday January 19, 2038 at 03:14:08 UTC, when the value over- (/home/sangeeth/soft/release2/src/) and performs a cd to the new
flows and becomes negative. An error is returned if the desired date top directory, which in my case, will be the second directory
and time exceeds this limit as shown here: (/home/sangeeth/soft/release2/src/show/).
Alternatively, one can pop a particular directory from the list of
# showdate -y 1000 directories by giving the directory ID (the ID is displayed beside a
showdate: required time exceeds system limit directory when using dirs -l -p -d) to the popd command:

The presence of whitespace characters in the formatting directive $ popd +1

needs to be escaped or enclosed in quotes (single/double). So, the
command to display the date and time 18 hours, 30 minutes ago in More options available for using the above commands can be
Year-Month-Day Hour:Minute:Second format would be: found by viewing the man pages.
—Sangeeth Keeriyadath
# showdate -h -18 -M -30 -f "%Y-%m-%d %H:%M:%S"

showdate cannot obtain the date and time by specifying a

weekly offset and by taking a positive or negative offset from
any datum, not just the current date and time. Even though
showdate has been tried and tested rigorously, it is not perfect.
And, if anyone encounters a bug or feels that redesigning the
algorithm, implementing coding shortcuts or efficiently using
system resources can improve the program, please contact me
by e-mail at [email protected].
showdate was designed for computing and displaying the date
and time in the past or future depending on the command-line
options, specified as an offset from the current date and time. The
next step would be to augment showdate to specify weeks and the
ability to change its datum.
—Sandeep Sahore

UU Push and Pop dirs

The dirs command, combined with pushd and popd, is very effective
for tracking users’ directory changes. Suppose you have to make some
changes to the files present in the following directories:

I /home/sangeeth/soft/release2/src/

I /home/sangeeth/soft/release2/src/show/

I /home/sangeeth/soft/release2/doc/

Instead of noting down the directories on paper, do the following:

$ pushd /home/sangeeth/soft/release2/src/
$ pushd /home/sangeeth/soft/release2/src/show/
$ pushd /home/sangeeth/soft/release2/doc/

To list all the directories, do the following:

w w w. l i n u x j o u r n a l . c o m november 2007 | 19
 COLUMNS
AT THE FORGE

Thinking about APIs

The historical evolution of Web sites into applications.

REUVEN M. LERNER People are surprising and unpredictable. In the computer increasingly standard part of Web-based applications, but
industry, you hear this in nearly every interview with the despite everyone’s use of the term API, that acronym
designer of a popular software package. For example, Perl means different things to different people.
originally was designed for network programming and text Starting next month, I’m going to look at the latest
processing. This made it an obvious choice for Web pro- batch of APIs that sites such as Facebook are offering. But
gramming, but Larry Wall certainly didn’t know or expect this month, I want to take a step back and consider the
that when he first wrote the language, years before the different types of APIs that software publishers offer. This
Web was invented. is useful if you intend to extend, use and work with those
Users of a software package almost always will push APIs. Web development is increasingly a matter of tying
its limits. That’s why some of the most successful and together existing functionality from other sites, and
popular programs are those that encourage users to go understanding these APIs can be quite useful.
beyond the author’s imagination. In the early days of the It’s also important for Web developers to understand
software industry, such add-ons and plugins didn’t exist, the nature of APIs. If you want to create the next
which meant that the only way to get new functionality Facebook or Google, you’re going to need to create more
was to lobby the authors and then hope the next release than a winning product. You’re going to need an ecosys-
would include the needed features. In the world of open- tem of developers and third-party software around your
source software, anyone is theoretically able to add new core product. One of the best ways to do this is to create
features, but between the generally small number of core and promote APIs, letting people use your application as
developers and the famously loud debates that sometimes a platform, rather than a standalone program. By looking
erupt, it can take surprisingly long to add a new feature. around and seeing what others have done, we can get a
(And although it is always possible to fork a project, this better sense of just what the possibilities are and how
has social and technical drawbacks that often outweigh we might use them.
the benefits.)
Some programs have a long tradition of encouraging Read-Only Protocols
add-ons. GNU Emacs is best known as a text editor, but it In the beginning, when Tim Berners-Lee invented the
comes with a full-fledged version of the Lisp programming Web, he imagined it as a read-write medium. But for
language. You can create just about anything you want in most people who used the Web during the first decade,
Emacs, and people have done so, including mail-reading it was a read-only medium. You could view Web sites with
programs, Web browsers and an extremely sophisticated your browser, fill out forms with your browser, and that
calendar/diary. Photoshop caught on among graphic was about it. There was no API for reading Web sites; if
designers not only because of its powerful image editing you wanted to read the content of a site programmatical-
features, but also because of the large number of plugins ly—for example, in order to create a searchable index of
that were developed (and sold) for the platform. Microsoft all Web content—you needed to create your own “spider”
Office, much as it might be reviled by many Linux and program, as well as teach it to take apart the HTML.
open-source advocates, became popular because of its This changed in the late 1990s, when a number of
built-in programming language (VBA), as much as for its developers (most prominently, but not exclusively, includ-
built-in features. And, of course, the Firefox browser ing Dave Winer) created RSS, standing either for really
wouldn’t be half as useful to me if it weren’t for the simple syndication or RDF site summary. In either case,
half-dozen plugins that I have added to my software. the idea was to create a machine-readable, frequently
So, users push software to the limits, and software updated summary of a site’s contents. By checking a
publishers have been able to make their offerings site’s RSS feed, you could learn whether there had been
more useful by making it possible to extend their any updates. More significant, RSS feeds were formatted
programs. How does this translate into an era of in a standard way, with standard tags, making it fairly
Web-based software? And, what does this mean to easy for programs to poll a feed and extract only the
us as Web developers? information they wanted.
The answer is the increasingly ubiquitous application Unfortunately, the term RSS became both the generic
programming interface, or API. If you want your Web site term for syndication and the name for several incompati-
to be taken seriously as a platform, and not just an appli- ble (but similar) formats for syndication. A separate group
cation, you need to offer an API that lets users create, of developers created the Atom protocol, which many
modify and extend your application. APIs are becoming an people believe is superior to all of the various RSS formats.

22 | november 2007 w w w. l i n u x j o u r n a l . c o m
 RSS and Atom are still popular today. The most retrieve it.
common use of these syndication feeds is for blog and This came in several different forms, all of which still
news updates, allowing users to keep track of which are used today. The first was XML-RPC, a simple RPC
sites have updated their content. But, RSS and Atom protocol that used HTTP to send an XML-encoded func-
can be used in other ways as well, providing a simple, tion invocation on a remote server. The server turned
reliable and machine-readable version of various types the XML-RPC request into a local function call and sent
of data from a Web site. If you are looking to broadcast the result of that call (or an error message) in an XML-
regularly updated data, RSS and Atom probably are encoded response. The good news was (and is) that
going to be a good starting point. XML-RPC is simple to understand and use, that there
For example, the well-known development company are implementations in many different languages, and
37signals provides an Atom feed of recent activity in its that they are generally compatible and interoperable.
Highrise contact management system. As helpful as it At the same time, XML-RPC was unable to handle
might be to look at your own feed, it would be even some of the more complex data types that people wanted
more helpful to aggregate multiple people’s feeds into a to use. Plus, it didn’t have the official seal of approval or
single viewer, allowing, for example, a manager to get complete standard that would have been necessary for it
a sense of what (and how much) employees are getting to enter the corporate arena. So, some of the original
done each day. XML-RPC developers created SOAP (originally known as
the Simple Object Access Protocol, but now an acronym
Read-Write Protocols that doesn’t expand). SOAP is more sophisticated and
The idea that a Web site could provide a regularly updated, complete than XML-RPC, but it had a great many issues
machine-parseable version of its content whetted the with compatibility, implementation and complexity. Today,
appetite of many developers for more. Many developers there are SOAP implementations for many languages, and
wanted a method to add and modify data, as well as it continues to be used in a variety of ways, despite some

Take a look inside the

Linux toolbox. ®

978-0-470-08291-1 978-0-470-08292-8 978-0-470-08293-5

Available wherever books are sold.

Wiley and the Wiley logo are registered trademarks of John Wiley & Sons, Inc. All other trademarks are the property of their respective owners.
 COLUMNS
AT THE FORGE

compatibility issues. Google also has exposed a number of its applications

But, at the same time that XML-RPC and SOAP were to read-write APIs. Rather than use one of the existing
being touted as the foundations for a new type of interac- protocols, Google uses a version of Atom for both
tive, machine-parseable Web, along came Roy Fielding, requests and responses, along with a data format it calls
who described the current state of affairs as unnecessarily GData. There are read-write APIs for a number of Google’s
complex. He proposed that instead of using XML in applications, including the calendar, Blogger and the
both the request and the response, we instead use spreadsheet program. Programmers no longer are limited
Representational State Transfer (REST). In other words, by the interface that Google provides to their spreadsheet
the URL should contain everything needed to describe data; they may create their own programs that use the
the request, without any additional XML markup or spreadsheet for storage and retrieval. (One slightly far-
payload. The response, of course, could be in XML or fetched example would be the creation of a distributed
any other format. database server that depended on Google’s spreadsheet
The idea of published Web services, procedures for locking and version control.)
invoked via HTTP and URLs that transferred data in XML Although new APIs of this sort constantly are being
and other formats, soon became widespread. Creating rolled out, the trend has seemed clear. Make the data
and using Web services became the biggest thing, with easily available and downloadable by the users, in a
every company talking about how it would take advantage variety of formats. And, make it possible for them to
of such Web services. Many standards were proposed for interact with your Web-based application either using
describing and finding Web services; for all I know, these your Web site or (alternatively) the command line or
standards still exist, but for the average programmer, they their own home-grown application.
don’t, and I’m not sure if and when they ever will.
Facebook
Read-Write APIs Facebook, the social networking site started by Mark
Given two read-only protocols and three read-write proto- Zuckerberg, has become an extremely popular application
cols, it was a matter of time before people started to on the Web. Facebook users can connect with friends, join
create applications that would take advantage of these. groups of like-minded individuals and send messages to
Amazon was one of the first companies to do so, opening others. Early in 2007, Facebook became popular among
up its catalog in a set of Web services now known as developers, as well as users, for creating a developer API
Amazon E-Commerce Services, or ECS. Amazon made its that goes far beyond the APIs I have described above. In a
entire catalog available via ECS, and it allowed program- nutshell, Facebook invited developers to create and deploy
mers to choose between SOAP and REST. Over the years, new applications that are seamlessly integrated into the
ECS has become an increasingly sophisticated and capable full Facebook experience.
system, making it possible to retrieve particular slices of Facebook isn’t the only site that lets you incorporate
Amazon’s catalog and pricing information. your own code into the site. However, the style and
But, retrieving information from Amazon is only half methods of this integration are deeper on Facebook than
the story: Amazon also makes it possible to manage a I have seen elsewhere. In the Facebook model, your Web
shopping cart via ECS and even has some facilities for application still resides on your server, but its output is
managing third-party products for sale. Amazon has made displayed inside the user’s Facebook page, alongside
a huge commitment to ECS, and a large community of other Facebook applications. This definitely is something
developers and third-party software vendors now exist new and exciting; I can’t think of any other Web sites
around this facility. By turning Amazon into a platform for that make it possible for an independent developer to
software development, rather than a simple on-line store, distribute code that integrates into the Web site. The fact
Amazon simultaneously has made a community of people that you can use whatever language and platform you
dependent on ECS and has created opportunities for the prefer, communicating with Facebook in a certain way,
creation of software applications that otherwise never marks the beginning of a new kind of API, one in which
would have been built. users can affect the Web service as seen by all users, not
eBay, Google and Yahoo! (among others) also have just one particular user. The only other site I can think
provided a number of APIs via Web services, which devel- of in this camp is Ning, Marc Andreessen’s build-your-
opers can use and access using various protocols. I’ve own-social-network site.
read reports, which I can’t confirm but that I’m willing to Moreover, Facebook has taken a page from Amazon
believe, claiming the majority of requests submitted to and eBay, telling developers that they can go wild, using
eBay’s servers are through its Web services. Given that the Facebook network for commercial as well as nonprofit
most eBay users are not technically sophisticated enough reasons. Google has had a long-standing policy of allowing
to create their own HTTP clients, we may assume there access to its maps, for example, but only for publicly
are a number of software development shops that see accessible Web sites and reasons. It remains to be seen
eBay as a distribution platform, much as others might whether Facebook’s API will continue to be free of charge
see Windows or Linux as their platform. and open to all.

24 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Something this sophisticated cannot use any one of And, although Amazon, Google and eBay have
the protocols that I mentioned above. Rather, Facebook demonstrated the importance and potential of a platform-
uses a combination of protocols and techniques to com- centric view, Facebook is pioneering the incorporation of
municate with your Web application, making it possible third-party applications. True, most Facebook applications
for your programs to display their output alongside other created to date are simple or trivial. But, we can expect
Facebook applications. Moreover, Facebook makes it possi- that these applications will become increasingly sophisti-
ble for your application to grab certain pieces of the user’s cated and useful over time. Facebook’s willingness to open
Facebook data, so even though your application doesn’t up to third-party developers is good for everyone—except
have access to the back-end Facebook database, it still can for competing sites, such as MySpace and LinkedIn, which
know (and display) something about the user’s friends. still appear to see themselves as standalone sites, rather
Your application even can send messages and notifications than platforms for new applications.
to the user’s friends, although Facebook has discovered This month, I explained why I find Facebook’s API to
that this can lead to spamming, so it remains to be seen be new and exciting. Next month, we’ll look at how you
exactly what happens on this front. can create your own Facebook applications. Even if you
aren’t interested in creating applications for Facebook,
Conclusion you owe it to yourself to see how the latest generation
Web sites used to be nothing more than an electronic of Web applications allow themselves to be modified,
method for publishing and reading basic information not just queried.I
encoded in HTML. But, Web sites evolved into applica-
tions, which spawned the first generation of APIs that Reuven M. Lerner, a longtime Web/database developer and consultant, is a PhD
made it possible to read and write your data. Facebook is candidate in learning sciences at Northwestern University, studying on-line
the first of the new generation of Web sites that look at learning communities. He recently returned (with his wife and three children) to
themselves as a platform more than an application. their home in Modi’in, Israel, after four years in the Chicago area.
 COLUMNS
COOKING WITH LINUX

Because Nothing Says

High Performance
MARCEL GAGNÉ
Like a Good Race
Got Linux? Got a good accelerated video card as well?
Then, get behind the wheel and drive!
Repeat what you just told me, François. No! Don’t
tell me that! Yes, François, I know I told you to repeat
what you told me, but I was sincerely hoping I was
mistaken and that I had heard something entirely dif-
ferent. Something like, “Patron, I lost the key to the
filing cabinet”, not, “I lost the key to the wine cellar.”
In my first restaurant job, François, this would have
been considered a firing offense. No, mon ami, I am
not going to fire you, but this is terrible! What shall
we serve our guests?
Non! I see them walking toward the restaurant
now. Quickly, François, think of something. Why you?
Because you locked up the wine cellar minutes before Figure 1. The Tux500 Indianapolis racecar—notice the logo on
opening time, then proceeded to lose the key. That’s the hood.
why. Quoi? Serve beer from my personal réfrigérateur
de bière? That is not a bad idea at all, mon ami. You In honor of this momentous event, the first race
live to serve another day, non? Don’t look so hurt, game on tonight’s menu, SuperTuxKart, features the very
François. We have guests. Quickly! same Tux at the wheel. SuperTuxKart, maintained by
Ah, welcome, everyone, to Chez Marcel, where fine Joerg Henrichs, is an updated and enhanced version of
Linux fare is served with the most excellent wines from the original TuxKart, created by Steve Baker. If you’ve
one of the world’s premier wine cellars—that is, except played the original, you’ll be impressed by the new,
tonight. It seems my faithful waiter has lost the keys to the hugely improved, SuperTuxKart. The game features slick
wine cellar. Never fear, mes amis, though I love that won- new graphics, new tracks, cool characters, racing adven-
drous liquid fruit of the grape, your Chef also enjoys a tures and more. You take the controls as one of several
good beer, be it lager or ale. So, tonight’s menu will be characters, including Tux, Penny (his friend), Mr. Ice
accompanied by selections from my own réfrigérateur de Block, Sushi the Octopus (Figure 2), Yeti and others. You
bière. Please, sit and make yourselves comfortable.
François can offer you several excellent selections, all
favorites of mine, including Alexander Keiths’ India Pale
Ale, Kilkenny Irish Cream Ale, Sleeman’s Original, Unibroue
Maudite and several others.
As you know, mes amis, this issue’s theme is high
performance, which we all know can refer only to rac-
ing and automobiles. If the words high performance
and Linux don’t immediately generate the same associ-
ation in your mind, you should know that in point of
fact, Linux and car racing go very well together. The
2007 Indianapolis 500 featured the first ever Linux-
sponsored car. The brainchild of Ken Starks, aka helios,
the Tux500 Project succeeded in its aim to raise
enough money to place Tux, the familiar Linux mascot
created by Larry Ewing, on the hood of an Indy car. For Figure 2. SuperTuxKart features some colorful characters,
more details, go to www.tux500.com. many with equally colorful names.

26 | november 2007 w w w. l i n u x j o u r n a l . c o m
 ",), +/),

8SWRFRUHVDQG
*%''55$0
8SWRFRUHVDQG
*%''55$0

-") %# (!*, ),' (  ,&# #&#.3

#( $/( !& ) ' #) ,#.3

&).$/54(/7-5#((!2$#/2%#/-054).'0/7%2#!."%7%$'%$).4/#/-0!#43%26%23
*œ}œÊˆ˜ÕÝÊÀ>ˆÃiÃÊ̅iÊL>ÀÊ>}>ˆ˜Ê܈̅Ê̅iÊ>‡˜iÜÊ{‡Ü>ÞÊ*iÀvœÀ“>˜Vi7>ÀiÊ£ÈnäÊ>˜`Ên‡Ü>ÞÊ*iÀvœÀ“>˜Vi7>ÀiÊxnÇäÊ܈̅Ê̅iʘiÜÊ+Õ>`‡
œÀiÊ
Ê
"«ÌiÀœ˜ÒÊ«ÀœViÃÜÀðÊ+Õ>`‡
œÀiÊ
Ê"«ÌiÀœ˜ÒÊ«ÀœViÃÜÀÃÊvi>ÌÕÀiʘ>̈ÛiʵÕ>`‡VœÀiÊÌiV…˜œœ}Þʜ˜Ê>ÊȘ}iÊ«ˆiViʜvÊȏˆVœ˜ÊÜÊ`>Ì>ÊV>˜Ê“œÛiÊ
LiÌÜii˜ÊVœÀiÃÊ>ÌÊvՏÊëii`ÊvœÀʓœÀiÊivwÊVˆi˜ÌÊ`>Ì>ÊÅ>Àˆ˜}°Ê
½ÃÊÀiۜṎœ˜>ÀÞÊˆÀiVÌÊ
œ˜˜iVÌÊ
ÀV…ˆÌiVÌÕÀiÊiˆ“ˆ˜>ÌiÃÊ̅iÊLœÌ̏i˜iVŽÃʈ˜…iÀi˜Ìʈ˜Ê
ÌÀ>`ˆÌˆœ˜>ÊvÀœ˜Ì‡Ãˆ`iÊLÕÃÊ>ÀV…ˆÌiVÌÕÀiÃ]ÊvœÀʅˆ}…‡Ì…ÀœÕ}…«ÕÌÊÀi뜘ÈÛi˜iÃÃÊ>˜`ÊÃV>>LˆˆÌÞÊ̜ʈ“«ÀœÛiʜÛiÀ>ÊÃÞÃÌi“ÊivwÊVˆi˜VÞ°Ê Vœ˜œ“ˆV>Êˆ˜ÊÈâiÊ
>˜`Ê«ÀˆViÊ>˜`ÊLՈÌÊvœÀÊ̅iʓœÃÌÊ`i“>˜`ˆ˜}Ê`>Ì>L>ÃiÊ>««ˆV>̈œ˜Ã]ʅˆ}…‡«iÀvœÀ“>˜ViÊVœ“«Ṏ˜}]Ê>˜`ÊÃiÀÛiÀÊۈÀÌÕ>ˆâ>̈œ˜]Ê̅iÊ*iÀvœÀ“>˜Vi7>ÀiÊ
£ÈnäÊ>˜`Ê*iÀvœÀ“>˜Vi7>ÀiÊxnÇäÊ`iˆÛiÀ°Ê
>Ê*œ}œÊ̜`>Þt

Ý«iÀˆi˜Vi]Ê“>}ˆ˜>̈œ˜]Ê>˜`Ê-Õ««œÀÌ°Ê
*œ}œÊˆ˜ÕÝʈÃÊ>À`Ü>ÀiÊ-œṎœ˜ÃÊ ÕˆÌÊvœÀʈ˜ÕÝ]Ê ÕˆÌÊvœÀÊ9œÕ°
ÜÜÜ°«œ}œˆ˜ÕÝ°Vœ“
/œÊ}iÌÊÃÌ>ÀÌi`]ÊVœ˜Ì>VÌÊÕÃÊ>ÌÊnnn°nÓn°*""ʜÀÊVœ˜Ì>VÌÊÃ>iÃJ«œ}œˆ˜ÕÝÊ
*œ}œÊˆ˜ÕÝ]ʘV°ÊÇä£Êˆv̅Ê
Ûi°Ê-ՈÌiÊÈnxä]Ê-i>Ì̏i]Ê7
ʙn£ä{

]Ê
̅œ˜]Ê
̅œ˜Ê8*]Ê>˜`Ê̅iÊ
ʏœ}œÊ>ÀiÊÀi}ˆÃÌiÀi`ÊÌÀ>`i“>ÀŽÃʜvÊ̅iÊ
`Û>˜Vi`ʈVÀœÊiۈViÃʜÀʈÌÃÊÃÕLÈ`ˆ>ÀˆiÃʈ˜Ê̅iÊ1˜ˆÌi`Ê-Ì>ÌiÃÊ>˜`ʜ̅iÀVœÕ˜ÌÀˆið
œÀÊ>``ˆÌˆœ˜>ÊÌiÀ“ÃÊ>˜`ÊVœ˜`ˆÌˆœ˜ÃÊ«i>ÃiÊۈÈÌÊÜÜÜ°«œ}œˆ˜ÕÝ°Vœ“
 COLUMNS
COOKING WITH LINUX

Figure 3. Explosions? They never told me there would Figure 4. On the Track with VDrift against a Computer Player
be explosions!
game called VDrift. Created by Joe Venzon, VDrift is an
can get SuperTuxKart from supertuxkart.berlios.de. exciting game based on drift racing. The V in VDrift refers
Your next step is to choose whether you want to to the Vamos Automotive Simulator, written by Sam
race on your own (for practice) or enter a race with Varner. Vamos is the physics engine that powers the game.
other players. These can be other humans (masquerad- Aside from being exciting and great fun to play, VDrift
ing as SuperTuxKart characters, of course), or they can comes with 19 world-class tracks and 28 different car
be computer AI players. Once your cast is assembled, models. VDrift is available at vdrift.net.
it’s time to select a race type and course. You can In case you don’t know what drift racing is (your
indulge in a Grand Prix event that takes you through humble Chef did not), it’s a form of racing where the
all 13 courses, a handful, or simply go for a single race driver employs “drifting” to control the car. Doesn’t
and select a time trial. help? Well, this quote from Wikipedia might clear
Once you are lined up at the starting line, don’t be things up: “A car is said to be drifting when the rear
too hasty putting the pedal to the metal or you’ll be slip angle is greater than the front slip angle and the
fined penalty points. Once you are off, your car can do front wheels are pointing in the opposite direction of
some interesting things besides drive. For starters, you the turn.” In other words, the car is turning one way,
can momentarily jump above your competition and any but the wheels are pointing in another direction. This,
obstacles in your path. Keep moving and collect spin- strange as it may seem, this is not an accident.
ning cubes along the way, as these contain things like Remember, the driver is in control. As the driver, what
rockets, which you then can fire at opponents and you are doing is sliding into a turn to avoid slowing
obstacles alike (Figure 3). Try to pick up the spinning down as much as possible.
fish coins as well—a penguin’s gotta eat. When the game starts, you can select the car you want
The courses in SuperTuxKart are fun, colorful and and, in some cases, the color. Where you race is another
imaginative. Race though an undersea lair, through the option. Choose a racetrack or go for a relaxing, high-speed
shifting sands of the Egyptian desert or around obstacles
in Oliver’s Math Glass. And, don’t worry too much about
falling off the edge of the world or an altogether
psychedelic cliff. Your jumping/flying car will carry you
back to safety using its combination magnetic levitation
system and butterfly wings.
The next item on tonight’s menu is a great racing

Note:
All these packages are available as source code bundles from their respective
Web sites. Before you start compiling (unless you want to, of course), check
your distribution’s software repositories for binary packages. In all cases, I had
no trouble finding packages. Figure 5. VDrift lets you select from several different cars,
colors and racing venues.

28 | november 2007 w w w. l i n u x j o u r n a l . c o m
 COLUMNS
COOKING WITH LINUX

for those of you who experience a kind of mania when

it comes to racecar driving. The game, aptly named
ManiaDrive, is an incredibly fast-paced game with rapid
turns, nerve-wracking jumps (Figure 6) and a driving,
rocking, soundtrack (ManiaDrive is actually a clone of
Nadéo Studio’s Trackmania game). ManiaDrive features a
training mode designed to prepare you for the real thing
and a set of complex tracks that can be played locally or
on-line with other players. You can get ManiaDrive from
maniadrive.raydium.org.
When you start ManiaDrive, make sure you go
through the Beginner story, which will guide you
through the various moves that are expected of you.
When you are ready for the big time, choose the Pro
story mode. This game is more an endurance and skill
test than a race. Sure, there’s a clock ticking and your
speed is tracked (on-line, against others, no less), but
races last only a short time, some less than a minute.
If you are the type that gets bored quickly, ManiaDrive
is for you.
Speaking of fast, is it possible that the time has
Figure 6. Yes, it is drive in the country. Practice races put you on the track with gone by so quickly? Mon Dieu! And yet, it does appear
exactly what it looks no other cars to worry about. Races pit you against an that closing time is almost upon us. Well, mes amis,
like. There is a huge opponent. On a single computer, you can play VDrift alone despite the apparent crisis that we apparently faced
gap in the road. or against a computer opponent, at which point, you also tonight, I dare say the beer selection was most refresh-
Hammer down, jump, can decide on your opponent’s car model and color (Figure ing. I can tell from your approving nods that you agree
keep it straight and 5). VDrift also can be run as a server, at which point, your with me. Nevertheless, tomorrow, I shall have the wine
keep your cool. friends can join in for a networked game. cellar lock changed, and next time, our award-winning
Several option controls exist to make the game cellar will once again be open. But for now, François
more realistic, and more difficult, if you so choose. For will pour you another beer that you may enjoy at your
instance, you can decide what effect speed has on leisure while you enjoy another race. Remember, mes
steering by adjusting a slider to a desired percentage. amis, that sitting in front of your Linux systems driving
The higher the percentage, the more difficult it is at a virtual car is the only way to drink and drive safely.
high speeds. You also can choose to map different keys Raise your glasses, mes amis, and let us all drink to one
for the gas pedal, brakes and steering. Shifting and another’s health. A votre santé! Bon appétit!I
clutch controls can be defined as well. I confess that I
found it much easier to set up to use an automatic Marcel Gagné is an award-winning writer living in Waterloo, Ontario. He is the
transmission on my first few races. That brings up author of the all-new Moving to Free Software, his sixth book from Addison-
another point—make sure you are in gear before you Wesley. He also makes regular television appearances as Call for Help’s Linux
guy. Marcel is also a pilot, a past Top-40 disc jockey, writes science fiction and
start. It took me a few tries before I realized that I was
fantasy, and folds a mean Origami T-Rex. He can be reached via e-mail at
going backward. Press the S key to start, then press 1 [email protected]. You can discover lots of other things (including great
to get in gear. If you have selected automatic, shifting Wine links) from his Web site at www.marcelgagne.com.
will happen automatically for you after this.
There are more options and settings, which I’ll let
you discover, but I do want to mention one other group Resources
of settings that may be important. Given that this an
OpenGL game and that it does require hardware 3-D ManiaDrive: maniadrive.raydium.org
acceleration, owners of somewhat less-powerful video
cards may find the game more responsive by toning SuperTuxKart: supertuxkart.berlios.de
down some of the effects.
As you race, you can change your viewing angle by VDrift: vdrift.net
pressing F1 through F5. If you find yourself totally taken
with the action on-screen and you feel the desire to pre- Marcel’s Web Site: www.marcelgagne.com
serve the moment, you can get a screenshot of the action
at any time by pressing F12. Those images will appear in The WFTL-LUG, Marcel’s Online Linux User Group:
the folder .vdrift/screenshots in your home directory. www.marcelgagne.com/wftllugform.html
The final item on tonight’s menu is the perfect selection

30 | november 2007 w w w. l i n u x j o u r n a l . c o m
 COLUMNS
WORK THE SHELL

Keeping Score in Yahtzee

Push an array sort out to the sort function instead of writing a sort routine.

DAVE TAYLOR Last month, I started talking about how to use some If we initialize the game by filling in all the array values
simple shell scripting techniques to create a computer sim- with a known stop value, say -1, then the test for whether
ulation of the popular dice game Yahtzee. I’m not going a given box has been filled in is easy:
to write the entire game (the computer player would be
darn complicated for a shell script, for one thing), but let’s if [ scoresheet[1] != -1 ] ; then
spend one more column looking at some of the basics of echo "1: Your one's" ; fi
scoring before we move on to other topics.
One nice thing about Yahtzee as a computer game is The trick is that we also want to pre-qualify these
that there really isn’t much work for the computer to do, options. There’s no point in prompting players to select
because the game works fine as a solitaire variation: you their roll as a Yahtzee if they didn’t get five of a kind. This
simply can play to try to maximize your score and see how proves to be a bit tricky, so as a first step, I’m going to
you do, competing against your previous high score. tweak the code to order the dice in the array in ascending
The basic idea of Yahtzee is that you roll five dice up order after each roll automatically.
to three times to get the best possible set of values. Those It might be a long trip to accomplish the task, but
are then scored against a fixed score pad that offers spe- rather than write a sorting routine in the script, I’m just
cific points for specific combinations, ranging from one going to push out the task to the sort function, then read
point per “one” in your roll up to 50 points for five of a the results back in and fit them into the individual slots of
kind (a “Yahtzee”, in game parlance). the dice array. Sound complicated? It is, rather:
A quick visit to Google identifies an on-line Yahtzee
score sheet; take a look at it before we proceed: function sortDice()
www.gamingcorner.nl/images/sheets/yahtzee-3.pdf. {
In many ways, it’s five-card draw poker with dice, so sorted="$( ( echo ${dice[1]} ; echo ${dice[2]}
the first section of the score sheet is for sets: ones, twos, echo ${dice[3]} ; echo ${dice[4]}
threes and so on, where your score for each is the sum of echo ${dice[5]} ) | sort )"
that value on the dice. The second section is other poker
hands, including three of a kind, four of a kind, small index=1
straight (four of the five dice have sequential values), large for value in $sorted ; do
straight (all five are sequential), a full house and one or dice[$index]=$value
more Yahtzee rolls. index=$(( $index + 1 ))
You can get bonuses for achieving certain levels, done
including a very nice 35-point bonus for attaining at }
least 63 on the top section, but the score sheet itself
is straightforward. You can see here that I’m using a temp variable called
The key to the game, then, is to figure out how to sorted to store the resultant values, and that I’m using a
score a given roll. If you roll four ones, do you want to subshell—that’s the $( ) notation—to do the actual work.
score that as your four of a kind or your ones? What if The hardest part of this little function is to figure out how
they’re both filled in already? Fortunately, we’re going to put the values back into the array once everything’s
to defer to the player for that, but that still leaves us sorted properly, and that’s accomplished with the for loop.
with the data question of how to model the different Notice that, by a lucky coincidence, for loops automat-
boxes on the score sheet and the interface question of ically step through fields separated by white space (spaces
how to prompt the user to select which box to score. and carriage returns), so it’s perfect for breaking the resul-
Let’s take a look. tant sorted sequence back into individual values.
We’re running low on space this month, and I’m afraid
Modeling the Score Sheet as an Array I’ve ended up spending quite a bit of time talking, rather than
As with most of these types of data structures, we’ll use coding. I’ll make it up to you, dear reader, next month!I
an array to model the score sheet. Count the boxes on the
score sheet, and you’ll see 13 boxes total, including the Dave Taylor is a 26-year veteran of UNIX, creator of The Elm Mail System, and
special Yahtzee box where you can roll—and get credit most recently author of both the best-selling Wicked Cool Shell Scripts and Teach
for—more than one (so a game typically has 13 rolls, but Yourself Unix in 24 Hours, among his 16 technical books. His main Web site is at
it could have more). www.intuitive.com, and he also offers up tech support at AskDaveTaylor.com.

32 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Intel Processor Blade w/ 3.5 in.
Hot-swappable Drive Bays

Much Better TCO than 1U Servers

AMD Quad Processor Blade

Chassis Management Module

Gigabit Ethernet Module

CMM Web-based GUI

4x DDR Infiniband switch

Application-Optimized for:
Enterprises, Financial Services, Databases, Data Centers
Science Labs, HPC, Personal Super Computer

www.supermicro.com CMM IPMI View

 COLUMNS
BEACHHEAD

Navigating by the Sun

Lowest TCO for sale to highest bidder.
“Dead reckoning”, cried Davi. “What is that?” Then, on June 27, 2007, I received another letter from Sun:
JON “MADDOG” HALL I had been adjusting the compass of my sailboat, the
Agape, measuring for possible magnetic influences that would A lot of myths have been circulating about Linux—
affect the readings, and I had mentioned the dead reckoning that it’s low cost; that it runs more apps; that it’s
method of navigation to Davi. Now, I knew I was going to pay pervasive...but when you look at the facts you see
the price of explanation. a different story.
“Dead reckoning is the method of determining where
you are by knowing where you have been, adding to that Red Hat 5.0 does not save you money, runs fewer cer-
the direction and speed you have been traveling, and the tified applications, and makes migration more difficult.
length of time you have been going in that direction to These are just three of the many compelling facts that
determine a new position. It was used for many hundreds should make you wonder, “why are you running Linux
of years at sea”, I told him. when you could be running Solaris 10 OS?”
“Wow, that must have been hard”, said my young friend.
“And it must have generated many errors.” Now, Total Cost of Ownership (TCO) studies abound
“Yes”, I answered, “and the method was blamed for the showing how Microsoft is cheaper than Sun, Sun is cheaper
loss of hundreds of ships and thousands of lives, including the than Linux, Linux is cheaper than Microsoft, and so on.
wreck of the HMS Association and several other ships from the Normally, these TCO studies show whatever the sponsoring
British fleet off the Isles of Scilly on October 22, 1707—1,400 company wants them to show, and it does not take heaps of
British sailors died.” intelligence to realize that a TCO study that does not show the
I continued, “This accident led John Harrison, a carpen- desired “winner” would not see the light of day. Although
ter, to develop the marine chronometer, a clock accurate TCO is important (because if you cannot afford the solution,
enough to calculate longitude easily while at sea. Harrison’s you cannot implement it), the really interesting indicator is
life is a fascinating story of perseverance and hard work, Return on Investment (ROI) and long-term benefits, which I call
which led to many important inventions, including bi-metal Value of Investment (VOI). VOI is usually what a company
strips and the roller bearing. Unfortunately, England’s Board wants for the long run.
of Longitude kept rejecting the work done by Harrison, citing Regarding the claim that Red Hat 5.0 “does not save you
inaccuracies in the running of the clocks, and it was felt that money”, Sun offered one example of a solution offered to
part of this rejection was because Harrison was a humble Marc Andreessen. Sun’s solution was cheaper than “Linux on
carpenter and not a gentleman. Eventually, the king had to Intel commodity hardware”. Okay, fine. But, this particular
intervene to see that justice was done.” customer also was Marc Andreessen, and having worked in
Aboard the Agape, e-mail is my prime method of commu- the UNIX Group at Digital Equipment Corporation (DEC) for
nication, and lately, I have been receiving messages from Sun 16 years, I can imagine the discussion between the various
Microsystems, telling me I should be switching from Linux to management groups inside Sun to make sure it won this
Solaris. They cite many “reasons” for doing this, and although account and to make sure it was “a good deal”, of which
some are features of Solaris that have not yet been incorporat- there is no detailed analysis of the terms listed in the e-mail
ed into Linux, most of them reek of marketing hype. or on the site. There is nothing wrong, illegal or even immoral
The first message I received (May 25, 2007) told me about this deal and its subsequent use in marketing (and
that “Solaris offers one thing Linux does—and 600 things certainly not on Marc’s part); it’s just a bit opportunistic by
it does not. The Solaris OS is free and open source—same Sun to use a well-known name on a single deal to make its
as Linux.” Yet, nowhere in the rest of the message does point—not scientific, by any means.
Sun identify those 600 things that Solaris offers. Nor could As for the claim that Solaris supports more “certified
I find them listed on Sun’s Web site where the links guided applications” than Red Hat v5.0, Sun used the application
me. Although the message does identify several features base of its SPARC and its x86 architectures combined. How
that Solaris has “built in”, some of those features are many people want to run both a SPARC and an Intel box
offered by major Linux distributions, and others can be just to get the complete suite of applications? Additionally,
added by “rolling your own” in the Linux space. the Solaris OS version had been out for a while, giving
Two of the major features Sun offers are “low-cost independent software vendors (ISVs) a chance to certify
licensing and support options” and “a disruption-free their applications. Red Hat’s v5.0 had been out for only a
migration path from the difficulties of a Linux platform to month or so at the time of the “data effectiveness” cited in
the ease and value of a Solaris 10 solution”. Hmm, sounds an e-mail message on April 4, 2007. This is hardly enough
a bit like marketing hype to me. time for the ISVs to certify their software on top of Red

34 | november 2007 w w w. l i n u x j o u r n a l . c o m
Hat’s newest release and and get it into Red Hat’s system. Again,
there’s nothing wrong or illegal, it’s just a little sleazy as a marketing
ploy—okay, more than a little sleazy.
Sun’s claim that Red Hat “makes migration more difficult”, brings about
the question “migration from what?” Last night, I went to my Linux user
group meeting, and the speaker, who was running a mixed Solaris and Linux
shop, lamented how hard it was to work on a Solaris system because of the
antiquated and nonstandard set of utilities on the upper levels of the operat-
ing system. His presentation confirmed some of the issues I knew about and
showed that being a Sun customer was not all peaches and cream. But then
again, I do not expect Sun’s marketing group to haul out the spoiled milk.
Sun also claims binary compatibility between releases, but when you
actually read Sun’s terms and conditions on binary compatibility, it basically
says Sun guarantees your program to run as long as you did not use any
interfaces that did not change or were not deleted. This is a guarantee? It
sounds like most of the warranties in the EULAs of most software products
I have seen—basically worthless.
I received a final e-mail message on June 28, 2007, inviting me to tell
Sun why I used Linux, and offering me a free 4GB iPod Nano if I came in
for a one-hour meeting to listen to why I should use Solaris. I would get
the iPod if I signed up and if I was qualified and if the number of iPods (25)
had not been depleted. Hmm, not that I even want an iPod, but what was
the chance of getting that iPod? That’s not illegal or immoral or even
sleazy—just marketing.
Why am I making a big deal about what obviously is a badly positioned
marketing campaign? After all, I was in the marketing group for DEC the
last eight years of my employment there. I know what it’s like in a large
company with many product groups and many marketing strategies that
often run in conflict with each other. I imagine some people inside Sun
really believe in FOSS and the community, and gnash their teeth whenever
they hear of such things. I even know one.
First, Sun purports to be a friend of the Free and Open Source
community and a reseller of Linux and Linux hardware. It has had four
of its server systems certified by at least one Linux distribution, and it
is listed as a Gold Partner of that Linux distribution. I assume that it is
making money doing this and perhaps even is selling services for those
servers and the Linux systems on them.
But, then I hear reports (and receive e-mail messages) that show this is
simply a bait-and-switch program—to get Linux customers onto its cus-
tomer list and convince them to switch to Solaris. With partners like this,
you really don’t need competitors.
Second, Sun has had a long and troubled history with maintaining the
consistency it brags about—corporate decisions that went in the face of
what its customer base really wanted, such as:

I The very painful migration of BSD SunOS to System 5.0-based Solaris,

and the needless renaming (and subsequent system administration
issues) of the BSD-based SunOS code base to Solaris 1.x (even though it
was the same code base).

I Supporting and then decommitting from Solaris Intel, not once,

but twice.

I Ignoring the fact that SPARC is big-endian and Intel is little-endian, so

that binary data often is incompatible. (“NFS takes care of that”, says
Sun at a tradeshow. “No, it does not”, says maddog back to them at
the same show.) Endianism was a point that Sun kept pushing as long
 COLUMNS
BEACHHEAD

as Big-Endian machines were prominent in the market- tactical solution, although the strategic (and long-term) solu-
place, but it has ignored this ever since Intel/AMD reached tion may still be FOSS. This even might lead to me advocating
dominance and SPARCs were the poor kids on the block. a Solaris solution for my customer. But this is what I, as the
solution provider, work out with my customer.
Third, Sun needs to decide what it wants to be when it On the opposite side, being a Linux Guy also means
grows up. It continues to spend millions of dollars to duplicate that when I engage Sun as a supplier of Linux servers for my
a base operating system, when it could be addressing cus- customers, I want to feel safe that I can turn my back for a
tomers’ needs by better supporting the Linux community and moment and not have the Sun people who service or resell
recognizing that the real enemy is a certain closed-source, those servers approach my customers and tell them that
proprietary operating system from a company that does not my solution is bad, particularly when that story is based on
allow real competition. Sun needs to redefine its strategy and half-baked marketing messages with no staying power.
become a solutions company (like IBM and HP are doing), When I was at DEC, I was making fairly good money in
while not abandoning its line of VARs and resellers that made a fairly stable job with good benefits. I saw no issue in sell-
Sun great. Sun needs to learn how to compete on the really ing both Digital UNIX and Linux on the Alpha. What my
equal playing field that computer products are now proceed- customers asked for, I sold them. When customers asked my
ing toward. It can do that, but only if it is smarter about it. opinion, I gave them the best answer I could for their solu-
Fourth, Sun needs to clean up its marketing messages. It tion. At that time, Digital UNIX was a commercially robust
claims to have been “founded as an open-source company”. and feature-full implementation, and Linux lacked in a lot of
In reality, Sun was one of the companies that took UNIX into a the features that Digital UNIX offered. I did not tell cus-
binary distribution model. Although it is true that this model tomers to buy Linux or buy Digital UNIX, because they both
was somewhat forced by AT&T’s arcane licensing strategies of made money for Digital. I could, however, see the writing
the time, it was this movement of UNIX companies to restrict on the wall. I knew that over time Linux would have all the
the source code distribution that helped motivate Richard features of a fully robust commercial system, that it would
Stallman to produce the GNU suite and the University of be used as the research arm for computer science, and that
California Berkeley to complete the first freely distributable companies pouring $200–300 million a year into developing
Berkeley Software Distribution. their “own version of UNIX” would not be able to compete
I do not want to imply that Sun has never helped the with an operating system and kernel being developed jointly
Open Source community or contributed to its well being. by many companies and many minds.
Certainly, through time there have been stellar examples of I left that stable job at DEC, because I believed in the Linux
individuals at Sun that have contributed more than their share community and the free software model, and because I
to the community overall. But, let’s put it into the perspective did not want to embarrass DEC by actively and aggressively
of a company that has had its share of proprietary, criticizing its biggest business partner, Microsoft.
closed-source products and standards wars. Now, Sun runs the risk of alienating what could be its
Sun does have some good products. Its work in multicore biggest and best business partner, the Linux community,
chips and multicore development tools (illustrated in an article including Linux VARs and Linux resellers.
this month by Roman Shaposhnik, a Sun employee, on page Just as ships no longer use marine chronometers and
80) is a good example of Sun leveraging off its technologies. sextants to “follow the Sun”, but use GPS systems instead,
Sun also has been doing some good work in environmentally companies no longer guide their businesses with half-baked
sound server systems and solutions. marketing statements.
Unfortunately, there is a big difference between having Sun’s marketing group needs to understand that this is no
good people working for you who believe in free and open- longer the time of sailing ships and dead reckoning. You can-
source software and what the corporate, management and not just start from where you have been and hope to end up
marketing culture states is the direction of the company. where you want by setting some course and sailing full steam
Here is my beef: I am a Linux Guy. I believe in free and ahead. You will find yourself on the rocks.
open-source software, and this means more than just the The product era of computer science is sinking, and now
open-source licenses that Sun offers. It means that customers the era of service-oriented computing is well underway. Sun’s
easily can contribute back to the path that Sun as a company marketing and business practices need to move into the 21st
takes, and it means that customers easily can contribute century, and Sun needs a new vision for how it is going to
back to the code that Sun writes. It means that as a business provide solutions for its customers and partners and move the
partner, I can feel free to propose my solutions to my company forward. I really hope Sun finds it.I
customers without having my partner run roughshod
over me. I have found my solution to free and open-source Jon “maddog” Hall is the Executive Director of Linux International (www.li.org), a
software in the Linux kernel and the people and companies nonprofit association of end users who wish to support and promote the Linux oper-
building and contributing to it. ating system. During his career in commercial computing, which started in 1969,
Mr Hall has been a programmer, systems designer, systems administrator, product
That said, when I engage customers, I engage them with
manager, technical marketing manager and educator. He has worked for such com-
the concept of solving their problems, and I am not above panies as Western Electric Corporation, Aetna Life and Casualty, Bell Laboratories,
using a less-open solution than what I might otherwise find if I Digital Equipment Corporation, VA Linux Systems and SGI. He is now an independent
think it is worthwhile for the customer’s VOI. I view this as a consultant in Free and Open Source Software (FOSS) Business and Technical issues.

36 | november 2007 w w w. l i n u x j o u r n a l . c o m
 61015C_71381202_IIALJXEO110007.indd
MAXIMIZE PROCESSING PERFORMANCE

InDesign CS2 / PDFx1a

AND MAXIMIZE RESPONSIVENESS.

06.18.07 12:30 AM
FILE NAME:

FORMAT:
DATE:
NOR-TECH
For information, contact Todd Swank
PORTABLE toll-free at 877-808-7438 or [email protected],
CLUSTER or visit www.nor-tech.com.

• Lightweight and
designed for mobility
Nor-Tech has developed a rugged and energy–efficient
• Ruggedized chassis Portable Cluster. The lightweight HPC clusters are available in
ensures safe transit shock-mounted 19" rackmount cases that adhere to military
• Ultra-energy efficient specifications and provide optimal environmental protection.

Intel, the Intel logo, Intel. Leap ahead., the Intel. Leap ahead. logo, Xeon, and Xeon Inside are trademarks of Intel Corporation in the U.S. and other countries. *Other names and brands may be claimed
as the property of others. © 2007 Intel Corporation. All rights reserved. Nor-Tech, the Nor-Tech logo, Nor-Tech HPC Clusters, and Nor-Tech Portable Clusters are trademarks or registered trademarks
of Northern Computer Technologies.

61015C_71381202_IIALJXEO110007.i1 1 6/19/07 12:48:59 PM

 COLUMNS
LINUX FOR SUITS

The Usefulness
Paradigm
DOC SEARLS If you don’t like the usefulness paradigm, submit a patch.

I don’t write code, but I do write prose. It’s different, 5. Do not join independent clauses by a comma.
but there are similarities and overlaps. Code writers have
also supplied me with a useful cache of metaphors. For 6. Do not break sentences in two.
example, I can explain my writing method as a debugging
and patching process—and not just because those are 7. A participial phrase at the beginning of a
metaphors for editing. Rather, it’s because I regard most of sentence must refer to the grammatical subject.
my writing less as finished work than as useful information
that is always subject to improvement. I’m not just here to 8. Divide words at line-ends, in accordance with
express. I’m here to help. (Okay, sometimes I just joke their formation and pronunciation.
around, but that’s different.)
My model is the most useful tome I’ve ever read, The III. Elementary Principles of Composition
Elements of Style, by William Strunk Jr. and E. B. White. Better
known just as “Strunk and White”, the book’s 1.0 version 9. Make the paragraph the unit of composition:
(en.wikisource.org/wiki/The_Elements_of_Style) was one paragraph to each topic.
written in 1918 by Strunk alone, when he was a professor
of English at Cornell University. It was privately published 10. As a rule, begin each paragraph with a topic
and only 43 pages long. It began: sentence; end it in conformity with the beginning.

This book is intended for use in English courses in 11. Use the active voice.
which the practice of composition is combined with
the study of literature. It aims to give in brief space 12. Put statements in positive form.
the principal requirements of plain English style. It
aims to lighten the task of instructor and student 13. Omit needless words.
by concentrating attention (in Chapters II and III) on
a few essentials, the rules of usage and principles 14. Avoid a succession of loose sentences.
of composition most commonly violated. The
numbers of the sections may be used as refer- 15. Express co-ordinate ideas in similar form.
ences in correcting manuscripts.
16. Keep related words together.
The book is a model of brevity and pith. The para-
graph above is first of the five that make up Chapter I, 17. In summaries, keep to one tense.
which fits on a single page. Here is Strunk’s outline for
Chapters II and III: 18. Place the emphatic words of a sentence at the end.

II. Elementary Rules of Usage Just reading that outline makes you a better writer.
Digging deeper has the same effect. Take #13 from
1. Form the possessive singular of nouns with ’s. Chapter III: “Omit needless words.” It begins:

2. In a series of three or more terms with a single Vigorous writing is concise. A sentence should
conjunction, use a comma after each term except contain no unnecessary words, a paragraph no
the last. unnecessary sentences, for the same reason that
a drawing should have no unnecessary lines and
3. Enclose parenthetic expressions between commas. a machine no unnecessary parts. This requires not
that the writer make all his sentences short, or
4. Place a comma before and or but introducing an that he avoid all detail and treat his subjects only
independent clause. in outline, but that every word tell.

38 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Familiar? Seems to me the same principle
applies to writing code. One difference is
that writers of prose generally need to be
instructed on the value of brevity while writ-
ers of code do not. That may seem a harsh
or inaccurate statement, but I make it
because code is purposeful in ways prose
does not have to be. In fact, blather is some-
times the very purpose of prose. Garrison
Keillor once called English “the preacher’s
language” because “it allows you to talk
until you think of what to say.”
I suppose you can blather in code too.
Back in the late 1990s, I was sitting in the
office of Phil Hughes, Linux Journal’s founder
and publisher. He had just come from an
argumentative session with a group of geeks
in another room. Imagining I would know
what he was talking about, he threw a pile of
code printout onto his desk. “Look at that!”,
he said. “Six pages of Perl! I could have done
the same thing in one page of awk!”
I’m not saying Perl is blather, by the way
(though Phil might). I am saying that omit-
ting needless words is a code-writing value
for which prose writers require help from the
likes of William Strunk.
Version 1.5 of Strunk and White came
with a 1935 revision by Edward A. Kenny
titled The Elements and Practice of
Composition. Strunk died in 1946. Far as I
know, his book didn’t change until 1957,
when the writer E. B. White, once a stu-
dent of Strunk’s, came again upon “the
little book” and set about improving it
with two short additional chapters and a
return to the original title. This first edi-
tion of what came to be called “Strunk
and White” was version 2.0 of what
began as Strunk’s alone. In the 86-page
1979 edition, White explained:

A second edition of the book was

published in 1972. I have now
completed a third revision. Chapter
IV has been refurbished with
words and expressions of a recent
vintage; four rules of usage have
been added to Chapter I. Fresh
examples have been added to
some of the rules and principles,
amplification has reared its head in
a few places in the text where I
felt an assault could successfully
be made on the bastions of its
brevity, and in general the book has
received a thorough overhaul—to
 COLUMNS
LINUX FOR SUITS

correct errors, delete bewhiskered entries, and centre of gravity is helping individuals make and
enliven the argument. implement better decisions at lower cost (where
“cost” includes time and hassle cost, as well
White died in 1985. The 1999 fourth edition of the as money cost). This is being made possible by
book has a forward by White’s stepson, Roger Angell, an information age: our increasing ability to
plus an index, a brief afterward by Charles Osgood and search for, gather, store, filter, edit, slice and
a glossary by Robert DiYanni. The current entry in dice, analyze, share the information we need
Wikipedia notes, “An anonymous editor modified the to for our decision-making.
text of this 1999 edition. Among other changes, he or
she removed White’s spirited defense of ’he’ for nouns Better personal decision-making (and implemen-
embracing both genders. See the ’they’ entry in tation) is an era-defining shift for two simple
Chapter IV and also gender-specific pronouns.” In his reasons: it shifts the focus of value creation
forward, Angell puts it more gently, “This edition has from companies and their goals to individuals
been modestly updated...with a light redistribution of and their goals, and it encompasses all aspects
genders to permit a feminine pronoun or female farmer of production, distribution and exchange (and a
to take their places among the males who once inno- whole lot more)....
cently served him. Sylvia Plath has knocked Keats out of
the box, and I notice that ’America’ has become ’this There is in this a declaration of independence for
country’ in a sample text, to forestall a subsequent and individuals, in heed of the need to equip human beings
possibly demeaning ’she’ in the same paragraph.” with better ways to learn, to make choices and to
Astute readers of Strunk, White and Linux Journal exercise their autonomy. What’s not clear, and needs
will observe that we violate one of the former pair’s to be, is the role played by programmers in equipping
original commands by putting our punctuation outside everybody with the freedom to escape the old persua-
our closing quotation marks, unless it is part of the sion paradigm, and a value system in which the depen-
quotation. This complies with a Linux Journal copy- dencies that matter most are ones where buyers rely
editing policy established by Phil Hughes in 1994, entirely on sellers.
in compliance with what had already become geek Free and open-source code does not only express
vernacular. One of my first arguments with Phil was the free choices of its authors, but it provides practical
over this very practice. I lost. ways to expand the freedoms of those who use and
Today if you look for The Elements of Style at depend on that code. Before I read Alan’s piece, I
Amazon.com, you’ll find three different books: The didn’t see the connection between proprietary code
1999 fourth edition, Strunk’s original edition and The and a larger paradigm anchored in persuasion. The
Elements of Style Illustrated, which (Wikipedia tells problem with proprietary code in many cases begins
us) is the fourth edition plus illustrations by Maira with motivation. It’s not just about making secret bits
Kalman. Somehow I can’t help regarding the three as and driving up sale value by making those bits scarce.
“distributions” of Strunk’s patched kernel, each useful It’s about controlling the customer. The persuasion
in its own way. paradigm puts a premium on captive, dependent
It’s also interesting to read and compare the discussion customers. Worse, it restricts the usefulness of prod-
and history pages behind the Wikipedia entry for The ucts to just what works for the seller. This usually
Elements of Style. The former calls to mind the Linux- involves limiting choices for customers.
Kernel Mailing List (LKML), while the latter serves as a kind What Alan calls the personal decision-making
of code repository. These analogies are far from precise, paradigm has been supported from Day Zero in the
but there is a common purpose in usefulness. Free and Open-Source Development world. What we
What brought me to this fresh appreciation of usefulness have here, I submit, is a usefulness paradigm—one
was a blog post by Alan Mitchell titled “Beyond the persuasion that has had a quiet and transformative influence on
paradigm” (rightsideup.blogs.com/my_weblog/2007/08/ all it supports. Today, that includes Linux, the Internet
beyond-the-pers.html). He writes: and countless other useful goods, all of which support
far more economic vitality and growth than we ever
Our modern commercial setup is organised enjoyed from the persuasion paradigm. (With due
around a persuasion paradigm, where the credit for all it did achieve in its day.)
driving force in commerce is companies and I share that observation with the intent that it will
their attempts to persuade individuals—so-called prove useful. If it’s not, we’ll patch it, discard it or rewrite
consumers—to buy their products or services it. If it is, any of us are still free to improve it.I
(and not those of their rivals).
Doc Searls is Senior Editor of Linux Journal. He is also a Visiting Scholar at the
We are, however, groping our way towards a University of California at Santa Barbara and a Fellow with the Berkman Center
personal decision-making paradigm, whose for Internet and Society at Harvard University.

40 | november 2007 w w w. l i n u x j o u r n a l . c o m
lj021:lj018.qxd 9/12/2007 5:40 PM Page 1

The Straight Talk People

SM

S I N C E 1 9 9 1
ABERDEEN
HERE IS HOW
THE BEST STACK UP
4-DRIVE 1U ABERNAS 8-DRIVE 2U ABERNAS
Up to 4TB Capacity Up to 8TB Capacity
• Intel® Pentium® 4 Processor • Single Dual-Core Intel Xeon® Processor
• 1GB DDR Memory • 2GB DDR2 Memory
• Available as Linux or Windows NAS • Available as Linux or Windows NAS
• 300W Power Supply • 500W Redundant Power
• From 1TB to 4TB • From 2TB to 8TB

Starting at
$
2,795 Starting at
$
4,295
12-DRIVE 2U ABERNAS 16-DRIVE 3U ABERNAS
Up to 12TB Capacity Up to 16TB Capacity
• Single Dual-Core Intel Xeon Processor • Dual Quad-Core Intel Xeon Processors
• 2GB DDR2 Memory • 2GB DDR2 Memory
• Available as Linux or Windows NAS • Windows Storage Server 2003
• 500W Redundant Power • 500W Redundant Power
• From 9TB to 12TB • From 8TB to 16TB

Starting at
$
8,495 Starting at
$
8,795
24-DRIVE 5U ABERNAS 32-DRIVE 6U ABERNAS
Up to 24TB Capacity Up to 32TB Capacity
• Dual Quad-Core Intel Xeon Processors • Dual Quad-Core Intel Xeon Processors
• 2GB DDR2 Memory • 2GB DDR2 Memory
• Windows Storage Server 2003 • Windows Storage Server 2003
• 950W Redundant Power • 1350W Redundant Power
• From 12TB to 24TB • From 16TB to 32TB

Starting at
$
11,795 Starting at
$
14,795
Features & Benefits:
• Up to 32TB of Storage Capacity
• Available in three Operating Systems
• GUI-Based Linux featuring iSCSI
• Windows Storage Server 2003 R2
• Windows Unified Data Storage
Server w/iSCSI
• Optimized SMB and NFS Performance
• Internal Windows OS Drive
• Linux OS on DOM
• Recovery from DVD or USB
• RAID 6 w/automatic rebuild
• Multi-platform File Sharing —
Windows, Linux, Unix, FTP, Macintosh

“Aberdeen surpasses HP … markedly higher scores …

AberNAS 128 boasts outstanding features”
Network Computing—Aberdeen AberNAS 128

Intel, Intel Logo, Intel Inside, Intel Inside Logo, Pentium, Xeon, and Xeon Inside are trademarks or registered trademarks of Intel Corporation or its
subsidiaries in the United States and other countries. For terms and conditions, please see www.aberdeeninc.com/abpoly/abterms.htm. lj021 888-297-7409
www.aberdeeninc.com/lj021
 NEW PRODUCTS

ASUS Eee PC 700/701

L
Fear not, gadget geeks, for in future issues we’ll fully cover the new ASUS Eee PC—so
much so that you’ll grumble “enough already!” The Eee PC is a new, ultraportable Linux-
based laptop starting at (this is not a typo) $259 US. ASUS is marketing the 700 model
(2GB Flash storage, 256MB of RAM) to first-time/elderly computer users, low-income house-
holds and K–12 education, while the 701 model (4GB Flash storage, 512MB of RAM) is
mainly for PC owners seeking convenient mobility. Both models feature preloaded
Xandros Linux, Intel Celeron-M 900MHz processor, 7" display, 10/100Mbps LAN,
802.11 b/g wireless, three USB 2.0 ports, MMC/SD card reader, VGA out,
Windows XP compatibility/drivers, built-in camera (optional on the 700)
and a four-cell battery. There are no optical drives; both models weigh
in at 0.9kg/2lbs. At the time of this writing, ASUS projects that
dealers should have the Eee PC by late September 2007.
www.asuseeepc.com

Lenovo’s ThinkPad
T Series Linux Notebooks
A year ago we saw Lenovo tepidly dip its pinky-toe into the Linux waters by certifying and supporting its
L

ThinkPad T60p notebook with user-installed SUSE Linux Enterprise Desktop (SLED). This year finds Lenovo
frolicking in the balmy Linux Sea, offering SLED preloaded and supported on its business-oriented
ThinkPad T Series. Lenovo says that strong demand for Linux notebooks from customers made the
preload option inevitable. Preloaded ThinkPads will be available to the general public in Q4 2007.
www.lenovo.com/linux

Storix, Inc.’s SBAdmin

L

Storix, Inc., recently released SBAdmin v6.2, touted as the first backup and system recov-
ery solution to integrate with IBM’s Tivoli Storage Manager (TSM). SBAdmin complements
TSM’s features and capabilities for Linux and AIX systems. SBAdmin also writes directly to
the TSM server, which cuts extraneous steps, saves time and increases reliability. In
addition, the application provides TSM users with disaster-recovery capabili-
ties, including a feature called Adaptable System Recovery, which enables
a system restore to the same or dissimilar hardware.
www.storix.com

Motorola’s MOTOMAGX
Mobile Linux Platform
In the near future, Motorola expects 60% of its handsets to run on Linux, and the new MOTOMAGX platform
is a means for realizing that goal. Motorola calls MOTOMAGX its “next-generation mobile Linux platform” that
will “deliver new levels of openness, flexibility and support for third-party applications” on its devices. The
company also hopes to empower its developer community to innovate in exciting ways. Currently, MOTOMAGX
supports apps developed in Java ME and will soon support WebUI and native Linux application environments.
www.motorola.com

42 | november 2007 w w w. l i n u x j o u r n a l . c o m
 NEW PRODUCTS

Olive’s OPUS 307S

Who knows where you’ll find Linux next? One
sure spot is the embedded brain inside Olive
Media Products’ new OPUS 307S audio system.
The OPUS 307S, part of Olive’s OPUS N°3
product line, is a stylish audio component that
goes in your stereo rack and contains all of your
digital music on its 250GB hard drive. The OPUS
can store up to 700 CDs in lossless quality.

L
Improvements on the OPUS 307S over previous
Olive devices include perpendicular recording
technology for lower power consumption, higher
storage density, improved long-term reliability
and inaudible operation. In addition, Olive will
preload up to 300 of your CDs for free.
www.olive.us
Lantronix’s
SecureLinx Branch
Office Manager

L
Move over Swiss Army knife, Lantronix has
released its new SecureLinx Branch (SLB)
Office Manager product, an IT management
appliance that integrates a console server, power
management and an Ethernet switch into a
1U rack-mountable device. With SLB, system
administrators can securely manage a diverse
range of servers and IT infrastructure equipment
in branch/remote offices from anywhere via the

Black Duck Internet. SLB allows enterprises to avoid the cost

of dedicated technicians and other expensive
Software’s protexIP overhead at each satellite location.
www.lantronix.com
Software development with open-source code is great, but it
can be complicated. This is why Black Duck Software created
protexIP/development (now v4.4), “a platform that helps compa-
nies govern how their software assets are created, managed
and licensed.” protexIP helps developers and legal counsel in
Sendio’s I.C.E Box E-Mail
managing the use of code from open-source projects who have
explicitly decided to switch to GPLv3 and those that have not. Also
Integrity Service Appliance
included is an enhanced KnowledgeBase, a library of open-source We’ve been tracking the buzz surrounding Sendio’s I.C.E Box, an
and vendor-added code software components with detailed e-mail security appliance that reputedly blocks 100% of spam with
licensing information for more than 140,000 components. no false positives. Although the I.C.E. Box has been shipping for
www.blackducksoftware.com several months, its Linux credentials recently came to our attention.

L Shunning the antispam filter approach, the I.C.E. Box performs a

one-time verification that the sender of the message is someone
with whom you indeed wish to communicate. The result is an end
to the junk while maintaining legitimate communications. The
appliance integrates seamlessly with any e-mail server and LDAP
environment. A built-in Kaspersky antivirus engine also is included.
www.sendio.com L

w w w. l i n u x j o u r n a l . c o m november 2007 | 43
 NEW PRODUCTS

Trinity’s Digital
Audio Workstation

L
Here is an audio device that is not brand-new but new to me and
hopefully to you too—Trinity Audio Group’s Trinity Digital Audio
Workstation. This slick, all-in-one device is a portable, professional
recording studio that allows one to do anything imaginable with linear
audio (for example, sample, edit, mix, play back and so on). Trinity
does everything your laptop with an audio interface can do, only more
conveniently and portably. Trinity runs on Linux and works with WAV,
MP3 and Ogg Vorbis files. Several core audio apps are included,
such as Audacity, Hydrogen drum machine, Ardour and more. Built-in
802.11g Wi-Fi lets you podcast from the field.
www.trinityaudiogroup.com

IBM’s Big Green

L
Linux Initiative
Thankfully, firms like IBM are starting to look at our serious environ-
mental problems as challenges rather than barriers. IBM’s new Big
Green Linux initiative seeks to leverage Linux and other technologies to
reduce costs and energy consumption by building cooler data centers
for itself and its customers. Big Green Linux is a subset of IBM’s
broader, yet similar, Project Big Green. Specifics of Big Green Linux
include server consolidation to System x and System p platforms; new
efficient products, such as the Information Server Blade; contributions
to the Linux kernel (such as tickless low-power state) and others.
www.ibm.com/linux

SugarCRM
L

By the time you flip to this page, SugarCRM will have its sweet new version 5.0
ready for you. This is the ninth major release of the popular commercial open-source
customer relationship management package. New features, such as the ability to
build custom modules, a new Ajax e-mail client and a Multi-Instance On-Demand
architecture, are said to “enhance the ability to build, customize, scale and upgrade
the application”. SugarCRM says its innovations are derived from feedback from
2,000 commercial customers and thousands of Open Source community members.
www.sugarcrm.com

Trusted Computer Solutions’ Security Blanket

Trusted Computer Solutions has long made security solutions for the CIA and other spy agencies in Washington, DC, and now its new
product, Security Blanket, aims to harden your Linux system with the same fervor. Trusted Computer Solutions calls Security Blanket
“a system lock-down and security management tool that enables systems administrators to configure and enhance the security level of
the Red Hat Enterprise Linux operating platform automatically”. The product’s advantage is to simplify the current arduous methods
(such as using Bastille) for “hardening” systems. Its menu-driven UI allows one to run either customized profiles or predefined ones
that automate industry-standard best practices from the National Institute of Standards and the Center for Internet Security. Security
Blanket supports Red Hat Enterprise Linux 4 and 5.
www.trustedcs.com

Please send information about releases of Linux-related products to James Gray at [email protected] or New Products
c/o Linux Journal, 1752 NW Market Street, #200, Seattle, WA 98107. Submissions are edited for length and content.

44 | november 2007 w w w. l i n u x j o u r n a l . c o m
Red Hat
Enterprise Linux
Cluster Suite
Building a highly available solution using the RHEL cluster suite. KHURRAM SHIRAZ

When mission-critical applications fail, so does your business. This Linux ES (for entry- and mid-range servers) and Red Hat Enterprise
often is a true statement in today’s environments, where most organi- Linux WS (for desktops/workstations). These products are designed
zations spend millions of dollars making their services available 24/7, specifically for use in enterprise environments to deliver superior appli-
365 days a year. Organizations, regardless of whether they are serving cation support, performance, availability and scalability.
external customers or internal customers, are deploying highly avail- The original release of Red Hat Enterprise Linux AS version 2.1
able solutions to make their applications highly available. included a high-availability clustering feature as part of the base prod-
In view of this growing demand, almost every IT vendor currently uct. This feature was not included in the smaller Red Hat Enterprise
is providing high-availability solutions for its specific platform. Famous Linux ES product. However, with the success of the Red Hat Enterprise
commercial high-availability solutions include IBM’s HACMP, Veritas’ Linux family, it became clear that high-availability clustering was a
Cluster Server and HP’s Serviceguard. feature that should be made available for both AS and ES server prod-
If you’re looking for a commercial high-availability solution on ucts. Consequently, with the release of Red Hat Enterprise Linux
Red Hat Enterprise Linux, the best choice probably is the Red Hat version 3 in October 2003, the high-availability clustering feature was
Cluster Suite. packaged into an optional layered product called the Red Hat Cluster
In early 2002, Red Hat introduced the first member of its Red Hat Suite, and it was certified for use on both the Enterprise Linux AS and
Enterprise Linux family of products, Red Hat Enterprise Linux AS (origi- Enterprise Linux ES products.
nally called Red Hat Linux Advanced Server). Since then, the family of The RHEL cluster suite is a separately licensed product and can be
products has grown steadily, and it now includes Red Hat Enterprise purchased from Red Hat on top of Red Hat’s base ES Linux license.

46 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Red Hat Cluster Suite Overview be run in standalone mode but introduces a single point of failure for
The Red Hat Cluster Suite has two major features. One is the GFS. Three or five GULM servers also can be run together, in which
Cluster Manager that provides high availability, and the other case the failure of one or two servers can be tolerated, respectively.
feature is called IP load balancing (originally called Piranha). GULM servers usually are run on dedicated machines, although this is
The Cluster Manager and IP load balancing are complementary not a strict requirement.
high-availability technologies that can be used separately or in In my cluster implementation, I used DLM, and it runs in each
combination, depending on application requirements. Both of cluster node. DLM is good choice for small clusters (up to two
these technologies are integrated in Red Hat’s Cluster Suite. In nodes), because it removes quorum requirements as imposed by
this article, I focus on the Cluster Manager. Table 1 shows the the GULM mechanism).
major components of the RHEL Cluster Manager. Based on DLM or GLM locking functionality, there are two basic
techniques that can be used by the RHEL
cluster for ensuring data integrity in concurrent
Table 1. RHEL Cluster Manager Components access environments. The traditional way is
the use of CLVM, which works well in most
SOFTWARE SUBSYSTEM COMPONENT PURPOSE RHEL cluster implementations with LVM-based
logical volumes.
Fence fenced Provides fencing infrastructure for specific Another technique is GFS. GFS is a cluster
hardware platforms. filesystem that allows a cluster of nodes to
access simultaneously a block device that is
DLM libdlm, Contains distributed lock management shared among the nodes. It employs distributed
dlm-kernel (DLM) library. metadata and multiple journals for optimal
operation in a cluster. To maintain filesystem
CMAN cman Contains the Cluster Manager (CMAN), integrity, GFS uses a lock manager (DLM or
which is used for managing cluster GULM) to coordinate I/O. When one node

x
membership, messaging and notification. changes data on a GFS filesystem, that change
is visible immediately to the other cluster nodes
GFS and related locks Lock_NoLock Contains shared filesystem support that can using that filesystem.
be mounted on multiple nodes concurrently. Hence, when you are implementing a RHEL
cluster with concurrent data access require-
GULM gulm Contains the GULM lock management ments (such as, in the case of an Oracle RAC

e
user-space tools and libraries (an alternative implementation), you can use either GFS or
to using CMAN and DLM). CLVM. In most Red Hat cluster implementa-
tions, GFS is used with a direct access configu-
Rgmanager clurgmgrd, clustat Manages cluster services and resources. ration to shared SAN from all cluster nodes.
However, for the same purpose, you also can
CCS ccsd, ccs_test Contains the cluster configuration services deploy GFS in a cluster that is connected to a
and ccs_tool dæmon (ccsd) and associated files. LAN with servers that use GNBD (Global
Network Block Device) or two iSCSI (Internet
Cluster Configuration System-config-cluster Contains the Cluster Configuration Tool, Small Computer System Interface) devices.
Tool used to configure the cluster and display Both GFS and CLVM use locks from the
the current status of the nodes, resources, lock manager. However, GFS uses locks from
fencing agents and cluster services graphically. the lock manager to synchronize access to
filesystem metadata (on shared storage), while
Magma magma and Contains an interface library for cluster lock CLVM uses locks from the lock manager to
magma-plugins management and required plugins. synchronize updates to LVM volumes and
volume groups (also on shared storage).
IDDEV iddev Contains the libraries used to identify the For nonconcurrent RHEL cluster implemen-
filesystem (or volume manager) in which a tations, you can rely on CLVM, or you can use
device is formatted. native RHEL journaling-based techniques (such
as ext2 and ext3). For nonconcurrent access
clusters, data integrity issues are minimal; I tried
to keep my cluster implementations simple by
Shared Storage and Data Integrity using native RHEL OS techniques.
Lock management is a common cluster infrastructure service that
provides a mechanism for other cluster infrastructure components to Fencing Infrastructure
synchronize their access to shared resources. In a Red Hat cluster, Fencing also is an important component of every RHEL-based cluster
DLM (Distributed Lock Manager) or, alternatively, GULM (Grand implementation. The main purpose of the fencing implementation is
Unified Lock Manager) are possible lock manager choices. GULM is a to ensure data integrity in a clustered environment.
server-based unified cluster/lock manager for GFS, GNBD and CLVM. In fact, to ensure data integrity, only one node can run a
It can be used in place of CMAN and DLM. A single GULM server can cluster service and access cluster service data at a time. The use

w w w. l i n u x j o u r n a l . c o m november 2007 | 47
FEATURE Red Hat Enterprise Linux Cluster Suite

of power switches in the cluster hardware configuration enables a Once downloaded, the package will be in tar format. Extract it,
node to power-cycle another node before restarting that node’s and then install at least the following RPMs, so that the RHEL cluster
cluster services during the failover process. This prevents any two with DLM can be installed and configured:
systems from simultaneously accessing the same data and corrupting
it. It is strongly recommended that fence devices (hardware or soft- VV Magma and magma-plugins
ware solutions that remotely power, shut down and reboot cluster
nodes) are used to guarantee data integrity under all failure condi- VV Perl-net-telnet
tions. Software-based watchdog timers are an alternative used to
ensure correct operation of cluster service failover; however, in most VV Rgmanager
RHEL cluster implementations, hardware fence devices are used, such
as HP ILO, APC power switches, IBM BladeCenter devices and the VV System-config-cluster
Bull NovaScale Platform Administration Processor (PAP) Interface.
Note that for RHEL cluster solutions with shared storage, an imple- VV DLM and dlm-kernel
mentation of the fence infrastructure is a mandatory requirement.
VV DLM-kernel-hugemem and SMP support for DLM
Step-by-Step Implementation
of a RHEL Cluster VV Iddev and ipvsadm
Implementation of RHEL clusters starts with the selection of proper
hardware and connectivity. In most implementations (without IP load VV Cman, cman-smp, cman-hugemem and cman-kernelheaders
balancing), shared storage is used with two, or more than two, servers
running the RHEL operating system and RHEL cluster suite. VV Ccs
A properly designed cluster, whether you are building a RHEL-
based cluster or an IBM HACMP-based cluster, should not contain any Restart both RHEL cluster nodes after installing vendor-related
single point of failure. Keeping this in mind, you have to remove any hardware support drivers and the RHEL cluster suite.
single point of failure from your cluster design. For this purpose, you
can place your servers physically in two separate racks with redundant Network Configuration
power supplies. You also have to remove any single point of failure For network configuration, the best way to proceed is to use the
from the network infrastructure used for the cluster. Ideally, you network configuration GUI. However, if you plan to use Ethernet
should have at least two network adapters on each cluster node, channel bonding, the configuration steps are slightly different.
and two network switches should be used for building the network Ethernet channel bonding allows for a fault-tolerant network
infrastructure for the cluster implementation. connection by combining two Ethernet devices into one virtual device.
The resulting channel-bonded interface ensures that if one Ethernet
Software Installation device fails, the other device will become active. Ideally, connections
Building a RHEL cluster starts with the installation of RHEL on two from these Ethernet devices should go to separate Ethernet switches
cluster nodes. My setup has two HP Proliant servers (DL740) with or hubs, so that the single point of failure is eliminated, even on the
shared fiber storage (HP MSA1000 storage). I started with a RHEL v4 Ethernet switch and hub level.
installation on both nodes. It’s best to install the latest available oper- To configure two network devices for channel bonding, perform
ating system version and its updates. I selected v4 update 4 (which the following on node 1:
was the latest version of RHEL when I was building that cluster). If you 1) Create bonding devices in /etc/modules.conf. For example,
have a valid software subscription from Red Hat, you can log in to the I used the following commands on each cluster node:
Red Hat network, and go to software channels to download the latest
update available. Later, once you download the ISO images, you can alias bond0 bonding
burn it to CDs using any appropriate software. During the RHEL OS options bonding miimon=100 mode=1
installation, you will go through various configuration selections, the
most important of which are the date and time-zone configuration, Doing this loads the bonding device with the bond0 interface name
the root user password setting, firewall settings and OS security level and passes options to the bonding driver to configure it as an active-
selection. Another important configuration option is network settings. backup master device for the enslaved network interfaces.
Configuration of these settings can be left for a later stage, especially 2) Edit the /etc/sysconfig/network-scripts/ifcfg-eth0 configuration
in building a high-availability solution with Ether-channel (or Ethernet file for eth0 and the /etc/sysconfig/network-scripts/ifcfg-eth1 file
bonding configuration). for the eth1 interface, so that these files show identical contents,
You may need to install additional drivers after you install as shown below:
the OS. In my case, I downloaded the RHEL support package for
the DL740 servers (the HP Proliant support pack, which is available DEVICE=ethx
from h18004.www1.hp.com/products/servers/linux/ USERCTL= no
dl740-drivers-cert.html). ONBOOT=yes
The next step is installing the cluster software package itself. This MASTER=bond0
package, again, is available from the RHEL network, and you definitely SLAVE=yes
have to select the latest available cluster package. I selected BOOTPROTO=none
rhel-cluster-2.4.0.1 for my setup, which was the latest cluster
suite available at the time. This enslaves ethX (replace X with the assigned number of the

48 | november 2007 w w w. l i n u x j o u r n a l . c o m
Ethernet devices) to the bond0 master device. Setup of the Fencing Device
3) Create a network script for the bonding device (for example, As I was using HP hardware, I relied on the configuration of the HP
/etc/sysconfig/network-scripts/ifcfg-bond0), which would appear like ILO devices as a fencing device for my cluster. However, you may
the following example: consider configuring other fencing devices, depending on the
hardware type used for your cluster configuration.
DEVICE=bond0 To configure HP ILO, you have to reboot your servers and press the F8
USERCTL=no key to enter into the ILO configuration menus. Basic configuration is rela-
ONBOOT=yes tively simple; you have to assign IP addresses to ILO devices with the name
BROADCAST=172.16.2.255 of the ILO device. I assigned 172.16.1.100 with Commilo1 as the name of
NETWORK=172.16.2.0 ILO device on node1, and 172.16.1.101 with Commilo2 as the ILO device
NETMASK=255.255.255.0 name on node2. Be sure, however, to connect Ethernet cables to the ILO
GATEWAY=172.16.2.1 adapters, which usually are marked clearly on the back side of HP servers.
IPADDR=172.16.2.182 Once rebooted, you can use the browsers on your Linux servers
to access ILO devices. The default user name is Administrator, with
4) Reboot the system for the changes to take effect. a password that usually is available on the hard-copy tag associated
5) Similarly, on node 2, repeat the same steps with the only differ- with the HP servers. Later, you can change the Administrator
ence being that the file /etc/sysconfig/network-scripts/ifcfg-bond0 password to a password of your choice, using the same Web-based
should contain an IPADDR entry with the value of 172.16.2.183. ILO administration interface.
As a result of these configuration steps, you will end up with
two RHEL cluster nodes with IP addresses of 172.16.2.182 and Setup of the Shared Storage Drive
172.16.2.183, which have been assigned to virtual Ethernet and Quorum Partitions
channels (the underlying two physical Ethernet adapters for each In my cluster setup environment, I used an HP fiber-based shared
Ethernet channel). storage MSA1000. I configured a RAID-1 of 73.5GB using the HP
Now, you easily can use the network configuration GUI on the smart array utility, and then assigned it to both of my cluster nodes
cluster nodes to set other network configuration details, such as using the selective host presentation feature.
hostname and primary/secondary DNS server configuration. I set After rebooting both nodes, I used HP fiber utilities, such
Commsvr1 and Commsvr2 as the hostnames for the cluster nodes and as hp_scan, so that both servers should be able to see this
also ensured that names resolution in both long names and short names array physically.
would work fine from both the DNS server and the /etc/hosts file. To verify the physical availability of shared storage for both cluster
A RHEL cluster, by default, uses /etc/hosts for node name nodes, look in the /dev/proc/proc file for an entry like /dev/sda or
resolution. The cluster node name needs to match the output of /dev/sdb, depending upon your environment.
uname -n or the value of HOSTNAME in /etc/sysconfig/network. Once you find your shared storage on the OS level, partition it
If you have an additional Ethernet interface in each cluster node, it according to your cluster storage requirements. I used the parted tool
on one of my cluster nodes to partition the shared storage. I created
two small primary partitions to hold raw devices, and a third primary
Listing 1. Contents of the /etc/hosts File on Each Server partition was created to hold the shared data filesystem:

# Do not remove the following line, or various programs Parted> select /dev/sda
# that require network functionality will fail.
127.0.0.1 localhost.localdomain localhost Parted > mklabel /dev/sda msdos
172.16.2.182 Commsvr1 Commsvr1.kmefic.com.kw
172.16.2.183 Commsvr2 Parted > mkpart primary ext3 0 20
172.16.1.186 Commilo1 Commilo1.kmefic.com.kw
172.16.1.187 Commilo2 Commilo2.kmefic.com.kw Parted > mkpart primary ext3 20 40
172.16.2.188 Commserver
192.168.10.1 node1 Parted > mkpart primary ext3 40 40000
192.168.10.2 node2
172.16.2.4 KMETSM I rebooted both cluster nodes and created the /etc/sysconfig/rawdevices
file with the following contents:

/dev/raw/raw1 /dev/sda1
is always a good idea to configure a separate IP network as an addi- /dev/raw/raw2 /dev/sda2
tional network for heartbeats between cluster nodes. It is important
that the RHEL cluster uses, by default, eth0 on the cluster nodes A restart of rawdevices services on both nodes will configure raw
for heartbeats. However, it is still possible to use other interfaces devices as quorum partitions:
for additional heartbeat exchanges.
For this type of configuration, you simply can use the network /home/root> services rawdevices restart
configuration GUI to assign IP addresses—for example, 192.168.10.1
and 192.168.10.2 on eth2, and get it resolved from the /etc/hosts file. I then created a JFS2 filesystem on the third primary partition using
the mke2jfs command; however, its related entry should not be put in

w w w. l i n u x j o u r n a l . c o m november 2007 | 49
FEATURE Red Hat Enterprise Linux Cluster Suite

It is strongly recommended use the same scripts for providing application/database monitoring and
high availability, so every application script used in a RHEL cluster
that fence devices (hardware should have a specific format.
or software solutions that All such scripts should at least have start and stop subsections, along
with a status subsection. When an application or database is available
remotely power, shut down and running, the status subsection of the script should return a value of
0, and when an application is not running or available, it should return a
and reboot cluster nodes) are value of 1. The script also should contain a restart subsection, which
used to guarantee data integrity tries to restart services if the application is found to be dead.
A RHEL cluster always tries to restart the application on the same
under all failure conditions. node that was the previous owner of the application, before trying to
move that application to the other cluster node. A sample application
the /etc/fstab file on either cluster node, as this shared filesystem will script, which was used in my RHEL cluster implementation (to provide
be under the control of the Rgmanager of the cluster suite:

/home/root> mke2jfs -j -b 4096 /dev/sda3 Listing 2. Sample Application Script

Now, you can create a directory structure called /shared/data on #Script Name: CommS.sh
both nodes and verify the accessibility of the shared filesystem from #Script Purpose: To provide application
both cluster nodes by mounting that filesystem one by one at each #start/stop/status under Cluster
cluster node (mount /dev/sda3 /shared/data). However, never try #Script Author: Khurram Shiraz
to mount this filesystem on both cluster nodes simultaneously, as it
might corrupt the filesystem itself. #!/bin/sh
basedir=/home/kmefic/KMEFIC/CommunicationServer
Cluster Configuration case $1 in
Almost everything required for cluster infrastructure has been done, so 'start')
the next step is configuring the cluster itself. cd $basedir
A RHEL cluster can be configured in many ways. However, the su kmefic -c "./CommunicationServer -f Dev-CommunicationServer.conf"
easiest way to configure a RHEL cluster is to use the RHEL GUI and exit 0
go to System Management→Cluster Management→Create a cluster. ;;
I created a cluster with the cluster name of Commcluster, and with 'stop')
node names of Commsvr1 and Commsvr2. I added fencing to both z=`ps -ef | grep Dev-CommunicationServer | grep -v "grep"|
nodes—fencing devices Commilo1 and Commilo2, respectively—so ¯awk ' { print $2 } '
that each node would have one fence level with one fence device. If `
you have multiple fence devices in your environment, you can add if [[ $? -eq 0 ]]
another fence level with more fence devices to each node. then
I also added a shared IP address of 172.16.2.188, which will be kill -9 $z
used as the service IP address for this cluster. This is the IP address that fuser -mk /home/kmefic
also should be used as the service IP address for applications or exit 0
databases (like for listener configuration, if you are going to use an fi
Oracle database in the cluster). ;;
I added a failover domain, namely Kmeficfailover, with priorities 'restart')
given in the following sequence: /root/CommunicationS.sh stop
sleep 2
Commsvr1 echo Now starting......
Commsvr2 /root/CommunicationS.sh start
echo "restarted"
I added a service called CommSvc and then put that service in the ;;
above-defined failover domain. The next step is adding resources to
this service. I added a private resource of the filesystem type, which 'status')
has the characteristic of device=/dev/sd3, mountpoint of /shared/data ps -U kmefic | grep CommunicationSe 1>/dev/null
and mount type of ext3. if [[ $? = 0 ]]
I also added a private resource of the script type (/root/CommS.sh) then
to service CommSvc. This script will start my C-based application, and exit 0
therefore, it has to be present in the /root directory on both cluster else
nodes. It is very important to have correct ownership of root and exit 1
security; otherwise, you can expect unpredictable behavior during fi
cluster startup and shutdown. ;;
Application or database startup and shutdown scripts are very esac
important for a RHEL-based cluster to function properly. RHEL clusters

50 | november 2007 w w w. l i n u x j o u r n a l . c o m
high availability to a legacy C-based application) is shown in Listing 2. You also have to pay special attention to keeping UIDs and GIDs
Finally, you have to add a shared IP address (172.16.2.188) to the synchronized across cluster nodes. This is important in making sure
service present in your failover domain, so that the service should contain proper permissions are maintained, especially with reference to the
three resources: two private resources (one filesystem and one script) and shared data filesystem.
one shared resource, which is the service IP address for the cluster. GRUB also needs to conform to the suite environment’s
The last step is synchronizing the cluster configuration across the specific needs. For instance, many system administrators, in a
cluster nodes. The RHEL cluster administration and configuration tool RHEL cluster environment, reduce the GRUB selection timeout
provides a “save configuration to cluster” option, which will appear to some lower values, such as two seconds, to accelerate system
once you start the cluster services. Hence, for the first synchronization, restart time.
it is better to send the cluster configuration file manually to all cluster
nodes. You easily can use the scp command to synchronize the Database Integration with a RHEL Cluster
/etc/cluster/cluster.conf file across the cluster nodes: The same RHEL cluster infrastructure can be used for providing
high availability to databases, such as Oracle, MySQL and IBM DB2.
/home/root> scp /etc/cluster/cluster.conf Commsvr2:/etc/cluster/cluster.conf The most important thing to remember is to base your
database-related services on a shared IP address—for example,
Once synchronized, you can start cluster services on both cluster nodes. you have to configure Oracle listener based on the shared service
You should start and stop RHEL-related cluster services, in sequence. IP address.
To start: Next, I explain, in simple steps, how to use an already-configured
RHEL cluster to provide high availability to a MySQL database server,
service ccsd start which is, no doubt, one of the most commonly used databases on RHEL.
service cman start I assume that the MySQL-related RPMs are installed on both
service fenced start cluster nodes and that the RHEL cluster already is configured with
service rgmanager start a service IP address of 172.16.2.188.
Now, you simply need to define a failover domain using the cluster
To stop: configuration tool (with the cluster node of your choice having a higher pri-
ority). This failover domain will have the MySQL service, which, in turn, will
service rgmanager stop have two private resources and one shared resource (the service IP address).
service fenced stop One of the private resources should be of the filesystem type (in my
service cman stop configuration, it has a mountpoint of /shared/mysqld), and the other
service ccsd stop private resource should be of the script type, pointing toward the
/etc/init.d/mysql.server script. The contents of this script, which should
If you use GFS, startup/shutdown of the gfs and clvmd services be available on both cluster nodes, is shown in Listing 3 on the LJ FTP
have to be included in this sequence. site at ftp.linuxjournal.com/pub/lj/listings/issue163/9759.tgz.
This script sets the data directory to /shared/mysqld/data, which is
Additional Considerations available on our shared RAID array and should be available from both
In my environment, I decided not to start cluster services at RHEL boot cluster nodes.
time and not to shut down these services automatically when shutting Testing for high availability of the MySQL database can be
down the RHEL box. However, if your business requires 24/7 service done easily with the help of any MySQL client. I used SQLyog,
availability, you can do this easily by using the chkconfig command. which is a Windows-based MySQL client. I connected to the
Another consideration is logging cluster messages in a different MySQL database on Commsvr1 and then crashed this cluster node
log file. By default, all cluster messages go into the RHEL log messages using the halt command. As a result of this system crash, the
file (/var/log/messages), which makes cluster troubleshooting somewhat RHEL cluster events were triggered, and the MySQL database
difficult in some scenarios. For this purpose, I edited the /etc/syslog.conf automatically restarted on Commsvr2. This whole failover process
file to enable the cluster to log events to a file that is different from the took one to two minutes and happened quite seamlessly.
default log file and added the following line:
Summary
daemon.* /var/log/cluster RHEL clustering technology provides a reliable high-available
infrastructure that can be used for meeting 24/7 business require-
To apply this change, I restarted syslogd with the service syslog restart ments for databases as well as legacy applications. The most
command. Another important step is to specify the time period for rotating important thing to remember is that it is best to plan carefully
cluster log files. This can be done by specifying the name of the cluster before the actual implementation and test your cluster and all
log file in the /etc/logrotate.conf file (the default is a weekly rotation): possible failover scenarios thoroughly before going live with a
RHEL cluster. A well-documented cluster test plan also can be
/var/log/messages /var/log/secure /var/log/maillog /var/log/spooler helpful in this regard.I
/var/log/boot.log /var/log/cron /var/log/cluster {
sharedscripts postrotate Khurram Shiraz is senior system administrator at KMEFIC, Kuwait. In his eight years of IT experience,
/bin/kill -HUP `cat /var/run/syslogd.pid 2> /dev/null` 2> he has worked mainly with IBM technologies and products, especially AIX, HACMP Clustering, Tivoli
/dev/null || true and IBM SAN/NAS storage. He also has worked with the IBM Integrated Technology Services group. His
endscript areas of expertise include design and implementation of high-availability and DR solutions based on
} pSeries, Linux and Windows infrastructures. He can be reached at [email protected].

w w w. l i n u x j o u r n a l . c o m november 2007 | 51
Getting Started
with Heartbeat
In every work environment with which I have been
Your first step toward involved, certain servers absolutely always must be up

high-availability bliss. and running for the business to keep functioning

smoothly. These servers provide services that always
need to be available—whether it be a database, DHCP,
Daniel Bartholomew DNS, file, Web, firewall or mail server.
A cornerstone of any service that always needs be
up with no downtime is being able to transfer the
service from one system to another gracefully. The
magic that makes this happen on Linux is a service
called Heartbeat. Heartbeat is the main product of
the High-Availability Linux Project.
Heartbeat is very flexible and powerful. In this
article, I touch on only basic active/passive clusters
with two members, where the active server is providing
the services and the passive server is waiting to take
over if necessary.

IMAGE ©ISTOCKPHOTO.COM/KATIV
52 | november 2007 w w w. l i n u x j o u r n a l . c o m
Installing Heartbeat The ha.cf File
Debian, Fedora, Gentoo, Mandriva, Red Flag, SUSE, Ubuntu and The next file to configure is the ha.cf file—the main Heartbeat config-
others have prebuilt packages in their repositories. Check your uration file. The contents of this file should be the same on all nodes
distribution’s main and supplemental repositories for a package with a couple of exceptions.
named heartbeat-2. Heartbeat ships with a detailed example file in the documenta-
After installing a prebuilt package, you may see a “Heartbeat tion directory that is well worth a look. Also, when creating your
failure” message. This is normal. After the Heartbeat package is ha.cf file, the order in which things appear matters. Don’t move
installed, the package manager is trying to start up the Heartbeat them around! Two different example ha.cf files are shown in
service. However, the service does not have a valid configuration Listings 2 and 3.
yet, so the service fails to start and prints the error message. The first thing you need to specify is the keepalive—the time
You can install Heartbeat manually too. To get the most recent between heartbeats in seconds. I generally like to have this set to
stable version, compiling from source may be necessary. There are one or two, but servers under heavy loads might not be able to
a few dependencies, so to prepare on my Ubuntu systems, I first send heartbeats in a timely manner. So, if you’re seeing a lot of
run the following command: warnings about late heartbeats, try increasing the keepalive.
The deadtime is next. This is the time to wait without hearing
sudo apt-get build-dep heartbeat-2 from a cluster member before the surviving members of the array
declare the problem host as being dead.
Check the Linux-HA Web site for the complete list of depen- Next comes the warntime. This setting determines how long to
dencies. With the dependencies out of the way, download the wait before issuing a “late heartbeat” warning.
latest source tarball and untar it. Use the ConfigureMe script to Sometimes, when all members of a cluster are booted at the
compile and install Heartbeat. This script makes educated guesses same time, there is a significant length of time between when
from looking at your environment as to how best to configure Heartbeat is started and before the network or serial interfaces are
and install Heartbeat. It also does everything with one command, ready to send and receive heartbeats. The optional initdead directive
like so: takes care of this issue by setting an initial deadtime that applies

sudo ./ConfigureMe install

Listing 2. The /etc/ha.d/ha.cf File on Briggs & Stratton
With any luck, you’ll walk away for a few minutes, and when
you return, Heartbeat will be compiled and installed on every node keepalive 2
in your cluster. deadtime 32
warntime 16
Configuring Heartbeat initdead 64
Heartbeat has three main configuration files: baud 19200
# On briggs the serial device is /dev/ttyS1
VV /etc/ha.d/authkeys # On stratton the serial device is /dev/ttyS0
serial /dev/ttyS1
VV /etc/ha.d/ha.cf auto_failback on
node briggs
VV /etc/ha.d/haresources node stratton
use_logd yes
The authkeys file must be owned by root and be chmod 600.
The actual format of the authkeys file is very simple; it’s only two
lines. There is an auth directive with an associated method ID
number, and there is a line that has the authentication method
and the key that go with the ID number of the auth directive. Listing 3. The /etc/ha.d/ha.cf File on Deimos & Phobos
There are three supported authentication methods: crc, md5
and sha1. Listing 1 shows an example. You can have more than keepalive 1
one authentication method ID, but this is useful only when you deadtime 10
are changing authentication methods or keys. Make the key warntime 5
long—it will improve security and you don’t have to type in the udpport 694
key ever again. # deimos' heartbeat ip address is 192.168.1.11
# phobos' heartbeat ip address is 192.168.1.21
ucast eth1 192.168.1.11
auto_failback off
Listing 1. The /etc/ha.d/authkeys File stonith_host deimos wti_nps ares.example.com erisIsTheKey
stonith_host phobos wti_nps ares.example.com erisIsTheKey
auth 1 node deimos
1 sha1 ThisIsAVeryLongAndBoringPassword node phobos
use_logd yes

w w w. l i n u x j o u r n a l . c o m november 2007 | 53
FEATURE Getting Started with Heartbeat

only when Heartbeat is first started. STONITH stands for Shoot The Other
You can send heartbeats over serial Node In The Head. STONITH devices are
or Ethernet links—either works fine. I
like serial for two server clusters that are
Sometimes, commonly some sort of network power-
control device. To see the full list of
physically close together, but Ethernet
works just as well. The configuration for
there’s only one supported STONITH device types, use the
stonith -L command, and use stonith
serial ports is easy; simply specify the
baud rate and then the serial device you
way to be sure -h to see how to configure them.
Next, in the ha.cf file, you need to list
are using. The serial device is one place
where the ha.cf files on each node may
whether a node your nodes. List each one on its own line,
like so:
differ due to the serial port having dif-
ferent names on each host. If you don’t
is dead, and node deimos
know the tty to which your serial port is
assigned, run the following command:
that is to kill it. node phobos

setserial -g /dev/ttyS*
This is where The name you use must match the output
of uname -n.

If anything in the output says “UART:

STONITH The last entry in my example ha.cf files is
to turn on logging:
unknown”, that device is not a real serial
port. If you have several serial ports, exper-
comes in. use_logd yes
iment to find out which is the correct one.
If you decide to use Ethernet, you have There are many other options that can’t
several choices of how to configure things. For simple two-server clusters, be touched on here. Check the documentation for details.
ucast (uni-cast) or bcast (broadcast) work well.
The format of the ucast line is: The haresources File
The third configuration file is the haresources file. Before configuring
ucast <device> <peer-ip-address> it, you need to do some housecleaning. Namely, all services that you
want Heartbeat to manage must be removed from the system init
Here is an example: for all init levels.
On Debian-style distributions, the command is:
ucast eth1 192.168.1.30
/usr/sbin/update-rc.d -f <service_name> remove
If I am using a crossover cable to connect two hosts together, I
just broadcast the heartbeat out of the appropriate interface. Here Check your distribution’s documentation for how to do the
is an example bcast line: same on your nodes.
Now, you can put the services into the haresources file. As
bcast eth3 with the other two configuration files for Heartbeat, this one
probably won’t be very large. Similar to the authkeys file, the
There is also a more complicated method called mcast. This haresources file must be exactly the same on every node. And,
method uses multicast to send out heartbeat messages. Check the like the ha.cf file, position is very important in this file. When
Heartbeat documentation for full details. control is transferred to a node, the resources listed in the
Now that we have Heartbeat transportation all sorted out, we haresources file are started left to right, and when control is
can define auto_failback. You can set auto_failback either to on or transfered to a different node, the resources are stopped right to
off. If set to on and the primary node fails, the secondary node left. Here’s the basic format:
will “failback” to its secondary standby state when the primary
node returns. If set to off, when the primary node comes back, it <node_name> <resource_1> <resource_2> <resource_3> . . .
will be the secondary.
It’s a toss-up as to which one to use. My thinking is that so The node_name is the node you want to be the primary on
long as the servers are identical, if my primary node fails, then the initial startup of the cluster, and if you turned on auto_failback,
secondary node becomes the primary, and when the prior primary this server always will become the primary node whenever it is up.
comes back, it becomes the secondary. However, if my secondary The node name must match the name of one of the nodes listed
server is not as powerful a machine as the primary, similar to how in the ha.cf file.
the spare tire in my car is not a “real” tire, I like the primary to Resources are scripts located either in /etc/ha.d/resource.d/ or
become the primary again as soon as it comes back. /etc/init.d/, and if you want to create your own resource scripts,
Moving on, when Heartbeat thinks a node is dead, that is just they should conform to LSB-style init scripts like those found in
a best guess. The “dead” server may still be up. In some cases, if /etc/init.d/. Some of the scripts in the resource.d folder can take
the “dead” server is still partially functional, the consequences are arguments, which you can pass using a :: on the resource line. For
disastrous to the other node members. Sometimes, there’s only example, the IPAddr script sets the cluster IP address, which you
one way to be sure whether a node is dead, and that is to kill it. specify like so:
This is where STONITH comes in.

54 | november 2007 w w w. l i n u x j o u r n a l . c o m
IPAddr::192.168.1.9/24/eth0 takes over the IP address, mounts the shared storage and
then starts up NFS. With this cluster, instead of having mainte-
In the above example, the IPAddr resource is told to set up a nance issues or other outages lasting several minutes to an
cluster IP address of 192.168.1.9 with a 24-bit subnet mask hour (or more), outages now don’t last beyond a second or two.
(255.255.255.0) and to bind it to eth0. You can pass other I can live with that.
options as well; check the example haresources file that ships
with Heartbeat for more information. Troubleshooting
Another common resource is Filesystem. This resource is for Now that your cluster is all configured, start it with:
mounting shared filesystems. Here is an example:
/etc/init.d/heartbeat start
Filesystem::/dev/etherd/e1.0::/opt/data::xfs
Things might work perfectly or not at all. Fortunately, with
The arguments to the Filesystem resource in the example logging enabled, troubleshooting is easy, because Heartbeat
above are, left to right, the device node (an ATA-over-Ethernet outputs informative log messages. Heartbeat even will let
drive in this case), a mountpoint (/opt/data) and the filesystem you know when a previous log message is not something you
type (xfs). have to worry about. When bringing a new cluster on-line, I
For regular init scripts in /etc/init.d/, simply enter them by usually open an SSH terminal to each cluster member and tail
name. As long as they can be started with start and stopped the messages file like so:
with stop, there is a good chance that they will work.
Listings 4 and 5 are haresources files for two of the clusters I run. tail -f /var/log/messages
They are paired with the ha.cf files in Listings 2 and 3, respectively.
The cluster defined in Listings 2 and 4 is very simple, and it Then, in separate terminals, I start up Heartbeat. If there are any
has only two resources—a cluster IP address and the Apache 2 problems, it is usually pretty easy to spot them.
Web server. I use this for my personal home Web server cluster. Heartbeat also comes with very good documentation. Whenever
The servers themselves are nothing special—an old PIII tower I run into problems, this documentation has been invaluable. On
and a cast-off laptop. The content on the servers is static HTML, my system, it is located under the /usr/share/doc/ directory.
and the content is kept in sync with an hourly rsync cron job.
I don’t trust either “server” very much, but with Heartbeat, I Conclusion
have never had an outage longer than half a second—not bad I’ve barely scratched the surface of Heartbeat’s capabilities here.
for two old castaways. Fortunately, a lot of resources exist to help you learn about
The cluster defined in Listings 3 and 5 is a bit more complicated. Heartbeat’s more-advanced features. These include active/passive
This is the NFS cluster I administer at work. This cluster utilizes and active/active clusters with N number of nodes, DRBD, the
shared storage in the form of a pair of Coraid SR1521 ATA-over- Cluster Resource Manager and more. Now that your feet are wet,
Ethernet drive arrays, two NFS appliances (also from Coraid) and hopefully you won’t be quite as intimidated as I was when I first
a STONITH device. STONITH is important for this cluster, because started learning about Heartbeat. Be careful though, or you might
in the event of a failure, I need to be sure that the other device end up like me and want to cluster everything.I
is really dead before mounting the shared storage on the other
node. There are five resources managed in this cluster, and to Daniel Bartholomew has been using computers since the early 1980s when his parents purchased
keep the line in haresources from getting too long to be readable, an Apple IIe. After stints on Mac and Windows machines, he discovered Linux in 1996 and has
I break it up with line-continuation slashes. If the primary cluster been using various distributions ever since. He lives with his wife and children in North Carolina.
member is having trouble, the secondary cluster kills the primary,

Listing 4. A Minimalist haresources File

Resources
stratton 192.168.1.41 apache2
The High-Availability Linux Project: www.linux-ha.org

Heartbeat Home Page: www.linux-ha.org/Heartbeat

Listing 5. A More Substantial haresources File Getting Started with Heartbeat Version 2:
www.linux-ha.org/GettingStartedV2
deimos \
IPaddr::192.168.12.1 \ An Introductory Heartbeat Screencast:
Filesystem::/dev/etherd/e1.0::/opt/storage::xfs \ linux-ha.org/Education/Newbie/InstallHeartbeatScreencast
killnfsd \
nfs-common \ The Linux-HA Mailing List:
nfs-kernel-server lists.linux-ha.org/mailman/listinfo/linux-ha

w w w. l i n u x j o u r n a l . c o m november 2007 | 55
 In early 2006,
Marshall University laid out a plan to
migrate HOBBIT (Figure 1), an HP
OpenVMS cluster handling university-wide
e-mail services. Plagued with increasing
spam attacks, this cluster experienced
severe performance degradation.
Although our employee e-mail store was
moved to Microsoft Exchange in recent
years, e-mail routing, mailing list and
Figure 1. HOBBIT OpenVMS Cluster Hardware Figure 2. Linux E-Mail Server Blades and SAN
student e-mail store (including IMAP
and POP3 services) were still served by
OpenVMS with about 30,000 active users.

Building
HOBBIT’s e-mail software, PMDF, provided
a rather limited feature set while charging
a high licensing fee. A major bottleneck
was discovered on its external disk storage

a Scalable
system: the dated storage technology
resulted in a limited disk I/O throughput
(40MB/second at maximal) in an e-mail
system doing intensive I/O operations.

High-Availability
To resolve the existing e-mail perfor-
mance issues, we conducted brainstorm-
ing sessions, requirements analysis, prod-
uct comparison and test-lab prototyping.
We then came up with the design of our

E-Mail System new e-mail system: it is named MUMAIL

(Figure 2) and uses standard open-source
software (Postfix, Cyrus-IMAP and MySQL)
installed on Red Hat Enterprise Linux. The

with Active core system consists of front-end e-mail

hub and back-end e-mail store. The
front-end e-mail hub uses two Dell blade
servers running Postfix on Linux. Network

Directory load balancing is configured to distribute

load between them. The back-end e-mail
store consists of two additional blade
servers running a Cyrus-IMAP aggregation

and More
setup. Each back-end node is then
attached to a different storage group on
the EMC Storage Area Network (SAN). A
fifth blade server is designated as a master
node to store centralized user e-mail
settings. Furthermore, we use LDAP and

A large-scale implementation Kerberos to integrate the e-mail user iden-

tities with Windows Active Directory (AD).

of a scalable Linux e-mail Figure 3 illustrates our new e-mail

system architecture and the subsystem

system with Active Directory. interactions with existing services, which

include Webmail, AD and SMTP gateway.
The block diagrams highlighted in red are
Jack Chongjie Xue the components to be studied in detail.

56 | november 2007 w w w. l i n u x j o u r n a l . c o m
Related Solutions
Before we zoom further into
our new e-mail system, I
want to mention some of the
existing Linux/UNIX e-mail
solutions in higher-education
environments. First, the HEC
Montréal e-mail system
discussed in a Linux Journal
article (see Resources) influ-
enced our design, which is
based on Cyrus-IMAP and
Postfix. Second, we looked
into Cambridge University’s
solution. It uses custom IMAP
proxy front-end servers and
multiple pairs of Cyrus-IMAP
mail store servers replicating
data to each other. Figure 3. System Architecture
Furthermore, Carnegie
Mellon University (CMU),
which originally developed Cyrus-IMAP, uses Sendmail as the including e-mail accounts, user e-mail routing, group aliases and
front-end mail exchanger and a Cyrus-IMAP Murder Aggregator mailing lists. Web-based user interfaces were developed using
setup on the back end. Columbia University moved its e-mail PHP to allow users to make changes to their settings in the
system to a Cyrus-IMAP-based solution in 2006, and the University database. Automated scripts running on the front-end nodes
of Indiana moved to Cyrus back in 2005. Cyrus and Postfix also will query the database for user settings and build Postfix maps
are used by Stanford University. to apply these settings.
Although the designs of these related solutions are different, A Postfix server can be thought of as routers (not for IP pack-
most of them use a cluster-based approach that separates ets but for e-mail). For each e-mail message, Postfix looks at the
mail transport/delivery from the mail store. Multiple front-end destination (envelope recipient) and the source (envelope sender)
MTA-MDA (Mail Transport Agent and Mail Delivery Agent) servers and then chooses how to route the e-mail message closer to its
are set up to deliver mail to the back-end mail store, which then destination. Lookup tables called Maps (such as Transport, Virtual,
saves messages either in a filesystem (for example, Maildir) or a Canonical and Alias Maps) are used to find the next-hop e-mail
database. Most of the solutions use Cyrus-IMAP (on UNIX or delivery location or apply e-mail address re-rewrites.
Linux) as their mail store server. A background job is running on each of the front-end e-mail
hub nodes to “pull” the e-mail settings (delivery location, e-mail
Detailed Design alias and group alias information) stored in the e-mail user database
Some distinctive differences set our design apart from the to the Postfix maps (aliases, virtual, canonical and transport).
existing solutions: Written in Perl, the program is configured to run periodically as
a crond job.
1. Instead of using a separate directory service (such as OpenLDAP) Our design principle of the new e-mail system is to scale out
for user authentication, our design integrates user identities with from a single, monolithic architecture to multiple nodes sharing
Windows Active Directory (AD). the same processing load. In a large e-mail environment, scaling
out the front-end MTA system is considerably easier compared
2. Rather than using an LDAP server to store user e-mail routing with scaling out the back-end mail store. As the front-end nodes
settings, we designed a relational database to store these settings. are essentially data-less, using DNS or IP-based load balancing on
multiple front-end servers is a typical practice. However, the same
3. In the mail store setup, instead of using an active-passive high- technique cannot be applied to design the back-end mail store
availability cluster setup, like the HEC approach or the Cyrus where the user data resides. Without clustering, shared storage or
replication approach developed at Cambridge, we deployed the additional software components (such as a proxy server), multiple
Cyrus-Murder Aggregator. Unlike the CMU Cyrus Aggregator mail store servers cannot share the same IMAP/POP3 process load
server allocation, which uses separate MTA server nodes, we under a unified service namespace. Because of this, using a single
consolidate both MTA and Cyrus Proxy functions to run on our mail store server tends to be an obvious solution. However, one
front-end mail hub nodes. node usually implies elevated server hardware expenses when
more powerful server hardware needs to be purchased to accom-
We designed an e-mail user database (running MySQL on the modate the ever-increasing system load. The price of a mid-range
Master node) to serve as a centralized data store for information server with four CPUs is usually much higher than the total price

w w w. l i n u x j o u r n a l . c o m november 2007 | 57
FEATURE Building a Scalable High-Availability E-Mail System with Active Directory & More

Figure 5. Linux Authentication and Authorization Against AD

the Cyrus Master node using the MUPDATE protocol, so that it

gets the information about which Cyrus back-end node stores
e-mail for the current client. Furthermore, the back-end Cyrus
nodes will notify the Master node about the mailbox changes
(creating, deleting and renaming mailboxes or IMAP folders) in
order to keep the master updated with the most current mailbox
location information. The Master node replicates these changes
to the front-end proxy nodes, which direct the incoming
IMAP/POP3/LMTP traffic. The MUPDATE protocol is used to
transmit mailbox location changes.
Although it is not a fully redundant solution (the Master node
is still a single point of failure), and half our users will suffer a
usage outage if either one of the back-end nodes is down, the
aggregator setup divides the IMAP processing load across multiple
servers with each taking 50% of the load. As a result of this
division of labor, the new mail store system is now scalable to
Figure 4. Cyrus-IMAP Aggregation Setup multiple servers and is capable of handling a growing user population
and increasing disk usage. More back-end Cyrus nodes can join
of three or more entry-class servers. Furthermore, a single-node with the aggregator to scale up the system.
architecture reduces system scalability and creates a single point
of failure. Integration with Active Directory
The Cyrus-IMAP package is proven to be robust and suitable in One of the requirements of our new e-mail system is to integrate
large settings. It differs from other Maildir or mbox IMAP servers user identities with the university directory service. Because Microsoft
in that it is intended to run as a “sealed” mailbox server—the Active Directory services have been made a standard within our
Cyrus mailbox database is stored in parts of the filesystem that are centralized campus IT environment, Cyrus (IMAP/POP3) and Postfix
private to the Cyrus-IMAP system. More important, a multiple (SMTP) are architected to obtain user authentication/authorization
server setup using Cyrus Murder aggregation is supported. It from AD. After the integration, all e-mail user credentials can be
scales out the system’s load by using multiple front-end IMAP managed from AD. Most directory services are constructed based
proxies to direct IMAP/POP3 traffic to multiple back-end mail store on LDAP. AD uses LDAP for authorization, and it has its own
nodes. Although we found other ways to scale out Cyrus-IMAP— Kerberos implementation for authentication. The goal of an
for example, Cambridge University’s pair-wise replication integrated AD authentication is to allow the Linux e-mail servers
approach, mentioned in the Related Solutions section of this to use AD to verify user credentials. The technology used to
article, or using a clustered filesystem to share IMAP storage support the AD integration scheme is based mainly on the
partitions between multiple servers with products like Red Hat’s Kerberos and LDAP support, which come with native Linux
Global File System (GFS)—compared with the aggregation components, as shown in Figure 5.
approach, these solutions either are too customized to support Here is how it works. First, we use AD Kerberos to authenti-
(the Cambridge approach) or involve extra cost (GFS is sold cate Linux clients. Pluggable Authentication Module (PAM) is con-
separately by Red Hat, Inc.). figured to get the user credentials and pass them to the pam_krb5
So, the Cyrus-IMAP Aggregation approach was adopted. library, which is then used to authenticate users using the Linux
Figure 4 illustrates the setup: two Cyrus back-end servers were set Kerberos client connection to the Key Distribution Center (KDC)
up, and each handles half the user population. Two Postfix MTA on Active Directory. This practice eliminates the need for authenti-
front-end nodes are designated to serve the proxy functions. cation administration on the Linux side. However, with only the
When e-mail clients connect through SMTP/IMAP/POP3 to the Kerberos integration, Linux has to store authorization data in
front-end servers, the Cyrus Proxy service will communicate with the local /etc/passwd file. To avoid managing a separate user

58 | november 2007 w w w. l i n u x j o u r n a l . c o m
authorization list, LDAP is used to retrieve user authorization informa- downtime throughout this period. All user e-mail messages
tion from AD. The idea is to let authorization requests processed by originally on HOBBIT were moved successfully to MUMAIL in a
Name Service Switch (NSS) first. NSS allows the replacement of many three-day migration window with automated and non-disruptive
UNIX/Linux configuration files (such as /etc/passwd, /etc/group migration processes. Users now experience significantly faster
and /etc/hosts) with a centralized database or databases, and the IMAP/POP3 access speed. Their e-mail storage quota is raised
mechanisms used to access those databases are configurable. NSS from 20MB to 200MB, and there is potential to increase the quota
then uses the Name Service Caching Dæmon (NSCD) to improve to a higher number (1GB). With the installation of gateway-level
query performance. (NSCD is a dæmon that provides a cache for spam/virus firewalls as well as increased hardware speed, no
the most common name service requests.) This can be very impor- e-mail backlog has been experienced on MUMAIL during recent
tant when used against a large AD user container. Finally, spam/virus outbreaks. With an Active Directory integrated user
NSS_LDAP is configured to serve as an LDAP client to connect to authentication setup, user passwords or other sensitive informa-
Active Directory to retrieve the authorization data from the AD tion are no longer stored on the e-mail system. This reduces user
users container. (NSS_LDAP, developed by PADL, is a set of C confusion and account administration overhead and increases
library extensions that allow LDAP directory servers to be used as network security. Mail store backup speed is improved significantly
a primary source of aliases, ethers, groups, hosts, networks, with faster disk access in the SAN environment. Finally, the new
protocol, users, RPCs, services and shadow passwords.) Now, system has provided a hardware and software environment that
with authorization and authentication completely integrated with supports future growth with the adoption of a scalable design.
AD using both LDAP and Kerberos, no local user credentials need More server nodes—both front end and back end—and storage
to be maintained. can be added when system usage grows in the future.I
In order to support LDAP authorization integration with Linux,
Windows Server 2003 Release 2 (R2), which includes support Jack Chongjie Xue holds a Masters’ degree in Information Systems from Marshall University,
for RFC 2307, is installed on each of the AD domain controllers. where he did Linux and Windows systems administration work. His job duties now include devel-
R2 introduces new LDAP attributes used to store UNIX or Linux oping Business Intelligence applications and working on data mining projects at Marshall.
user and group information. Without an extended AD LDAP
schema, like the one used by R2, the Linux automatic authoriza-
tion integration with AD is not possible. It is also important to Resources
mention that the SASL Authentication layer shown in Figure 3
is using Cyrus-SASL, which is distributed as a standard package “Migration of Alcator C-Mod Computer Infrastructure to Linux” by
by Carnegie Mellon University. The actual setup uses PAM for T. W. Fredian, M. Greenwald and J. A. Stillerman:
authenticating IMAP/POP3 users. It requires the use of a special www.psfc.mit.edu/~g/papers/fed04.pdf
Cyrus dæmon, saslauthd, which the SASL mechanism uses to
communicate via a Linux-named socket. “HEC Montréal: Deployment of a Large-Scale Mail Installation” by
Ludovic Marcotte: www.linuxjournal.com/article/9323
Conclusion
Our new e-mail system is mostly based on open-source software. Cyrus-IMAP Aggregation: cyrusimap.web.cmu.edu/ag.html
The incorporation of Postfix, Cyrus-IMAP and MySQL helped to
fulfill most of the system requirements. From the hardware per- “Scaling up Cambridge University’s E-Mail Service” by David Carter
spective, the technologies used, such as Storage Area Network and Tony Finch: www-uxsup.csx.cam.ac.uk/~fanf2/hermes/
(SAN), blade server and the Intel x86_64 CPUs, helped meet the doc/talks/2004-02-ukuug/paper.html
requirements of fast access, system scalability and high availability.
However, the use of open-source software and new hardware CMU’s Cyrus-IMAP Configuration:
technologies may introduce new management overhead. Although cyrusimap.web.cmu.edu/configuration.html
all the open-source software packages used on the new system
are mature products, compared with commercial software, they Columbia’s Cyrus-IMAP Move:
typically lack a GUI for system management. Their configuration www.columbia.edu/cu/news/05/12/cyrus.html
and customization are completely based on a set of plain-text
configuration files. Initially, this may present a learning curve, as Indiana’s Cyrus-IMAP Information:
the syntax of these configuration files must be studied. But, once uitspress.iu.edu/040505_cyrus.html
the learning curve is passed, future management easily can be
automated, as scripts can be written to manage the configuration Stanford’s E-Mail System Discussion:
parameters and store them in a centralized location. On the www.stanford.edu/dept/its/vision/email.html
hardware side, complex settings also may imply complex network
and server management settings, which also may introduce overhead Windows Security and Directory Services for UNIX Guide:
during system management. However, the benefits of using the www.microsoft.com/downloads/details.aspx?familyid=144f
technologies discussed outweigh the complexities and learning 7b82-65cf-4105-b60c-44515299797d&displaylang=en
curves involved. It is easy to overcome the drawbacks through proper
design, configuration management and system automation. “Toward an Automated Vulnerability Comparison of Open-Source
At the time of this writing, our new Linux e-mail system IMAP Servers” by Chaos Golubitsky: www.usenix.org/events/
(MUMAIL) has been running in production for ten months. The lisa05/tech/golubitsky/golubitsky.pdf
entire system has been running in a stable state with minimal

w w w. l i n u x j o u r n a l . c o m november 2007 | 59
DISTRIBUTED COMPILING
with distcc
You don’t need a cluster to get cluster-like
performance out of your compiler.
One of the most frustrating aspects of open-source development
is all the time spent waiting for code to compile. Right now,
compiling KDE’s basic modules and libraries on a single machine
takes me around three hours, and that’s just to get a desktop.
Even with a core 2 duo, it’s a lot of time to sit around and wait.
With another pair of core duo machines at my disposal, I’d
love to be able to use all of their processing power combined.
Enter distcc.
distcc is a program that allows one to distribute the load of
compiling across multiple machines over the network. It’s essen-
tially a front end to GCC that works for C, C++, Objective C and
Objective C++ code. It doesn’t require a large cluster of compile
hosts to be useful—significant compile time decreases can be
seen by merely adding one other similarly powered machine. It’s
a very powerful tool in a workplace or university environment
where you have a lot of similar workstations at your disposal, but
one of my favourite uses of distcc is to be able to do develop-
ment work on my laptop from the comfort of the café downstairs
and push all the compiles up over wireless to my more powerful
desktop PC upstairs. Not only does it get done more quickly, but
also the laptop stays cooler.
It’s not necessary to use the same distribution on each system,
but it’s strongly recommended that you use the same version of
GCC. Unless you have set up cross-compilers, it’s also required
that you use the same CPU architecture and the same operating
system. For example, Linux (using ELF binaries) and some BSDs
(using a.out) are not, by default, able to compile for each other.
Code can miscompile in many creative and frustrating ways if the
compilers are mismatched.

JES HALL

60 | november 2007 w w w. l i n u x j o u r n a l . c o m
Installation The order of the hosts is extremely important. distcc is unable to
The latest version of distcc, at the time of this writing, is 2.18.3. There determine which hosts are more powerful or under less load and
are packages for most major distributions, or you can download the simply distributes the compile jobs in order. For jobs that can’t be
tarball and compile it. It follows the usual automake procedure of run in parallel, such as configure tests, this means the first host in
./configure; make; make install; see the README and INSTALL the list will bear the brunt of the compiling. If you have machines of
files for details. varying power, it can make a large difference in compile time to put
distcc needs to be called in place of the compiler. You simply can the most powerful machines first and the least powerful machine
export CC=distcc for the compilers you want to replace with it, but last on the list of hosts.
on a development workstation, I prefer something a little more perma- Depending on the power of the computer running distcc, you may
nent. I like to create symlinks in ~/bin, and set it to be at the front of not want to include localhost in the list of hosts at all. Localhost has to
my PATH variable. Then, distcc always is called. This approach used to do all of the preprocessing—a deliberate design choice that means
work around some bugs in the version of ld that was used in building you don’t need to ensure that you have the same set of libraries and
KDE, and it is considered to have the widest compatibility (see the header files on each machine—and also all of the linking, which is
distcc man page for more information): often hugely processor-intensive on a large compile. There is also a
certain small amount of processing overhead in managing shipping
mkdir ~/bin for i in cc c++ gcc g++; do ln -s `which distcc` ~/bin/$i; done the files around the network to the other compilers. As a rule of
thumb, the distcc documentation recommends that for three to
If ~/bin is not already at the beginning of your path, add it to your four hosts, localhost probably should be placed last on the list,
shellrc file: and for greater than five hosts, it should be excluded altogether.
Now that you have your cluster configured, compiling is very simi-
export PATH=~/bin:$PATH lar to how you would have done it without distcc. The only real differ-
setenv PATH ~/bin:$PATH ence is that when issuing the make command, you need to specify
multiple jobs, so that the other machines in the cluster have some
for bourne- and C-compatible shells, respectively. work to do. As a general guide, the number of jobs should be approx-
imately twice the number of CPUs available. So, for a setup with three
Client Configuration single-core machines, you would use make -j6. For three dual-core
Each client needs to run the distcc dæmon and needs to allow machines, you would use make -j 12. If you have removed localhost
connections from the master host on the distcc port (3632). The from your list of hosts, don’t include its CPU or CPUs in this reckoning.
dæmon can be started manually at boot time by adding it to
rc.local or bootmisc.sh (depending on the distribution) or even
from an inetd. If distccd is started as an unprivileged user account,
it will retain ownership by that UID. If it is started as root, it will
attempt to change to the distcc or nobody user. If you want to
start the dæmon as root (perhaps from an init script) and change
to a user that is not distcc or nobody, the option -user allows you
to select which user the dæmon should run as:

distccd -user jes -allow 192.168.80.0/24

In this example, I also use the -allow option. This accepts a

hostmask in common CIDR notation and restricts distcc access to
the hosts specified. Here, I restrict access only to servers on the
particular subnet I’m using on my home network—machines with
addresses in the 192.168.80.1–192.168.80.254 range. If you are Figure 1. distccmon-text Monitoring a Compile
particularly security-conscious, you could restrict it to a single address
(192.168.80.5) or any range of addresses supported by this notation. distcc includes two monitoring tools that can be used to watch
I like to leave it pretty loose, because I often change which host is the the progress of compile jobs. The console-based distccmon-text is
master depending on what I’m compiling and when. particularly excellent if your master host is being accessed via SSH.
As the user the compile job is running as, execute the command
Compiling distccmon-text $s, where $s is the number of seconds at which
Back on the master system on which you plan to run your compiles, you would like it to refresh. For example, the following:
you need to let distcc know where the rest of your cluster is. There are
two ways of achieving this. You can add the hostnames or IP addresses distccmon-text 5
of your cluster to the file ~/.distcc/hosts, or you can export the variable
DISTCC_HOSTS delimited by whitespace. These names need to updates your monitor every five seconds with compile job information.
resolve—either add the names you want to use to /etc/hosts, or use The graphical distccmon-gnome is distributed as part of distcc if
the IP addresses of the hosts if you don’t have internal DNS: you compile from source, but it may be a separate package depending
on your distribution. It provides similar information in a graphical dis-
192.168.80.128 192.168.80.129 localhost play that allows you to see at a glance which hosts are being heavily
utilised and whether jobs are being distributed properly. It often takes

w w w. l i n u x j o u r n a l . c o m november 2007 | 61
FEATURE Distributed Compiling with distcc

It doesn’t require a large cluster of compile over SSH when needed.

hosts to be useful—significant compile time First, create an SSH key
using ssh-keygen -t dsa,
decreases can be seen by merely adding then add it to the target user’s
~/.ssh/authorized_keys on your
one other similarly powered machine. distcc hosts. It’s recommended
always to set a passphrase on
an SSH key for security.
In this example, I’m using
a few tries to get the order of hosts at the most optimal—tools like my own user account on all of the hosts and a simple bash loop to
distccmon-gnome make it easier to see whether machines are being distribute the key quickly:
under- or over-utilised and require moving in the build order.
for i in 192.168.80.120 192.168.80.100; do cat ~/.ssh/id_dsa.pub
¯| ssh jes@$i 'cat - >> ~/.ssh/authorized_keys'; done

To let distcc know that it needs to connect to the hosts under SSH,
modify either the ~/.distcc/hosts file or $DISTCC_HOSTS variable. To
instruct distcc to use SSH, simply add an @ to the beginning of the
hostname. If you need to use a different user name on any of the
hosts, you can specify it as user@host:

localhost @192.168.80.100 @192.168.80.120

Because I’m using a key with a passphrase, I also need to start

my SSH agent with ssh-add and enter my passphrase. For those
unfamiliar with ssh-agent, it’s a tool that ships with OpenSSH that
facilitates needing to enter the passphrase for your key only once
Figure 2. Graphical distcc Monitoring a session, retaining it in memory.
Now that we’ve set up SSH keys and told distcc to use a secure
Security connection, the procedure is the same as before—simply make -jn.
distcc relies on the network being trusted. Anyone who is able to con-
nect to the machine on the distcc port can run arbitrary commands on Other Options
that host as the distcc user. It is vitally important that distccd processes This method of modifying the hostname with the options you want
are not run as root but are run as the distcc or nobody user. It’s also distcc to honour can be used for more than specifying the connection
extremely important to think carefully about your -allow statements type. For example, the option /limit can be used to override the
and ensure that they are locked down appropriately for your network. default number of jobs that will be sent to the distccd servers. The
In an environment, such as a home or small workplace network, original limit is four jobs per host except localhost, which is sent only
where you’re security firewalled off from the outside world and two. This could be increased for servers with more than two CPUs.
people can be made accountable for not being good neighbours Another option is to use lzo compression for either TCP or SSH
on the network, distcc is secure enough. It’s extremely unlikely connections. This increases CPU overhead, but it may be worthwhile
that anyone could or would exploit your distccd hosts if they’re on slow networks. Combining these two options would be done with:
not running the dæmon as root and your allow statements limit
the connecting machines appropriately. localhost 192.168.80.100/6,lzo
There is also the issue that those on the network can see your
distcc traffic—source code and object files on the wire for anyone to This option increases the jobs sent to 192.168.80.100 to six, and it
reach out and examine. Again, on a trusted network, this is unlikely to enables use of lzo compression. These options are parsed in a specific
be a problem, but there are situations in which you would not want order, so some study of the man page is recommended if you intend
this to happen, or could not allow this to happen, depending on what to use them. A full list of options with examples can be found on the
code you are compiling and under what terms. distcc man page.
On a more hostile network, such as a large university campus or a The flexibility of distcc covers far more than explained here. One
workplace where you know there is a problem with security, these popular configuration is to use it with ccache, a compiler cache. distcc
could become serious issues. also can be used with crossdev to cross-compile for different architec-
For these situations, distcc can be run over SSH. This ensures both tures in a distributed fashion. Now, your old SPARC workstation can
authentication and signing on each end, and it also ensures that the get in on the act, or your G5 Mac-turned-Linux box can join the party.
code is encrypted in transit. SSH is typically around 25% slower due to These are topics for future articles though; for now, I’m going to go
SSH encryption overhead. The configuration is very similar, but it play with my freshly compiled desktop environment.I
requires the use of ssh-keys. Either passphraseless keys or an ssh-agent
must be used, as you will be unable to supply a password for distcc to Jes Hall is a UNIX systems consultant and KDE developer from New Zealand. She’s passionate
use. For SSH connections, distccd must be installed on the clients, but about helping open-source software bring life-changing information and tools to those who
it must not be listening for connections—the dæmons will be started would otherwise not have them.

62 | november 2007 w w w. l i n u x j o u r n a l . c o m
Hear Yourself Think Again!

WhisperStation™
Cool... Fast... Silent!
For 64-bit HPC, Gaming and Graphic Design Applications
Originally designed for a group of power hungry, demanding engineers in the automotive industry,
WhisperStation™ incorporates two dual core AMD Opteron™ or Intel® EM64T™ processors, ultra-quiet
fans and power supplies, plus internal sound-proofing that produce a powerful, but silent, computational
platform. The WhisperStation™ comes standard with 2 GB high speed memory, an NVIDIA e-GeForce
or Quadro PCI Express graphics adapter, and 20" LCD display. It can be configured to your exact
hardware specification with any Linux distribution. RAID is also available. WhisperStation™ will also
make a system administrator very happy, when used as a master node for a Microway cluster!
Visit www.microway.com for more technical information.
Experience the “Sound of Silence”.
Call our technical sales team at 508-746-7341 and design your personalized WhisperStation™ today.
 INDEPTH
Picking the RapidMind
RapidMind has a mind to make advances in the use of
multicore programming rapidly available. NICHOLAS PETRELEY

Writing applications to support multiple CPU cores is not an LJ: How much, if anything, does the programmer need to
easy task, and in some cases, it is even harder if you want to take know about parallel processing programming techniques
a huge existing application and adapt it for multiple cores. So I in order to use RapidMind?
figured the real breakthrough is likely to be years away. It seems McCool: We believe that developers are the application experts
as if RapidMind has a solution for this problem that doesn’t and should have some involvement in moving their applications
require a massive overhaul of an existing application, and its solution into the parallel world. The key is to let developers leverage what
is already available. they already know, rather than force them down an unfamiliar
We invited RapidMind’s President and CEO Ray DePaul and and frustrating path. RapidMind is built upon concepts already
Founder and Chief Scientist Michael McCool to talk about RapidMind’s familiar to all developers: arrays and functions. It is not necessary
approach to exploiting the power of multicore systems. for a developer to work directly with threads, vectorization, cores
We deemed it important to look at RapidMind, because it or synchronization. Fundamentally, a developer can apply func-
seems as if we’re finally entering the age of parallel processing on tions to arrays, and this automatically invokes parallel execution.
the desktop as chip manufacturers bump up against the practical A RapidMind-enabled program is a single-threaded sequence of
limits of Moore’s Law. Everything from graphics cards to PlayStation parallel operations and is much easier to understand, code and
3 consoles exploit parallel processing these days. I have an Intel test than the multithreaded model of parallel programming.
quad-core processor in my workstation. Although I’m happy with
it, I find that the only time I truly appreciate having this multicore LJ: Can you give us a simple code example (the includes and
chip is when I run multiple applications simultaneously or run declaration statements that would start a typical program)?
multiple processes, such as with the command make -j 5. If McCool: First, you include the platform header file and optionally
anything, single-threaded applications run slower on this chip activate the RapidMind namespace:
than on the single-core CPU I used to run, because each core in
the Intel chip is significantly slower (2GHz vs. 3GHz). #include <rapidmind/platform.hpp> using namespace rapidmind;
So how does RapidMind bridge the gap between existing software
and the changing computational model? Next, you can declare variables using RapidMind types for numbers
and arrays:
LJ: Could you give us a brief description of RapidMind, and the
problem it is designed to solve? Value1f f; Array<2,Value3f> a, b;
DePaul: RapidMind is a multicore software platform that allows
software organizations to leverage the performance of multicore The Value1f type is basically equivalent to a float, and the Array
processors and accelerators to gain a real competitive advantage types are used to manage large collections of data. These can be
in their industry. With RapidMind, you can develop parallel applica- declared anywhere you would normally declare C++ variables: as
tions with minimal impact on your development lifecycle, costs and members of classes or as local or global variables.
timelines. And, we allow you to accomplish this without the need A Program object is the RapidMind representation of a function
for multithreading. You leverage existing skills, existing compilers and is created by enclosing a sequence of operations on RapidMind
and IDEs and take advantage of all key multicore architectures types between RM_BEGIN and RM_END. The operations will then
without constantly porting your application. be stored in the Program object. For example, suppose we want to
add a value f, represented using a global variable, to every element
LJ: So is it accurate to say RapidMind is actually a library of of an array. We would create a program object prog as follows:
common C/C++ operations, where the exploitation of multiple
cores is largely transparent to the programmer? Program prog = RM_BEGIN {
McCool: RapidMind is much more than a simple library of In<Value1f> c; Out<Value1f> d;
“canned functions”. In fact, it is possible to use the API to the d = c + f;
RapidMind platform to specify an arbitrary computation, and for } RM_END;
that computation to execute in parallel with a very high level
of performance. We provide a sophisticated multicore software Note that although the program may run on a co-processor,
platform that can leverage many levels of parallelization, but at we can just refer to external values like f in the same way we
the same time allows developers to express their own computa- would from a function definition. It is not necessary to write
tions in a very familiar, single-threaded way. any other code to set up the communication between the host

64 | november 2007 w w w. l i n u x j o u r n a l . c o m
 RapidMind’s Founder and Chief Scientist Michael
McCool (left) and President and CEO Ray DePaul (right)

processor and any co-processors. ing, such as deadlocks or other synchronization issues?
To apply this operation to array a and put the result in array b, invok- McCool: The semantics of the RapidMind interface does not involve
ing a parallel computation, we just use the program object like a function: explicit locking or synchronization by the developer. The platform itself
automatically takes care of these issues when necessary at a lower
b = prog(a); level in the runtime platform. The developer cannot specify programs
that deadlock or that have race conditions, in the same way that a
Of course, in real applications, program objects can contain a Java developer cannot specify programs that have memory leaks.
large number of operations, and a sequence of program objects
and collective operations on arrays (such as scatter, gather and LJ: I see Hewlett-Packard software ran 32.2 times faster after
reduce) would be used. the software was adapted to use RapidMind. How long did it
take to modify the software to use RapidMind?
LJ: How do you avoid the common pitfalls of parallel process- McCool: Our collaboration with HP was a great test of our plat-

w w w. l i n u x j o u r n a l . c o m november 2007 | 65
 INDEPTH

form. Roughly the same amount of time was processor-specific optimization, data manage-
taken to RapidMind-enable the application as
was taken by HP to tune its single-core baseline
“ Foryouthecan
same effort,
use
ment, dynamic load balancing, scaling for
additional cores and multiple levels of paral-
version. The tuning by HP sped up its version lelization. The RapidMind platform performs
by a factor of 4, whereas RapidMind running RapidMind not only all of these functions.
on an NVIDIA 7900 GPU outperformed that by
a factor of more than 32. More recently, we
to get significantly LJ: You support multiple platforms on differ-
have run the same code on an NVIDIA 8800 higher performance ent levels. For example, you can exploit the
GPU and sped it up by an additional factor of processors on NVIDIA and ATI graphics cards,
5, and we also have run the RapidMind version on the same multicore the Cell processor, as well as multicore CPUs.
on our multicore CPU quad-core product and processors you’re In addition, you support both Linux and
achieved a speedup of 8 over HP’s version. Windows, correct?
So the benefit to the software organization already targeting, DePaul: The processor vendors are delivering
is quite startling. For the same effort, you can some exciting and disruptive innovations.
use RapidMind not only to get significantly
but you can leverage Software companies are faced with some
higher performance on the same multicore the additional tough choices—which vendors and which
processors you’re already targeting, but you architectures should they support. By leverag-
can leverage the additional performance of performance of ing RapidMind, they get to benefit from all of
accelerators as well. The RapidMind version the hardware innovations and deliver better
also will scale automatically to future processors
with more cores.
accelerators as well.
” products to their customers within their
current development cycles and timelines.
RapidMind will continue to provide portable
LJ: Is the speed increase in the HP software typical or “best performance across a range of both processors and operating systems.
case”? What software is most likely to see speed increases? We will support future multicore and many-core processors, so
Database server software? Complex queries on data ware- applications written with RapidMind today are future-proofed and
housing? Spam filtering? Web browsers? Something else? can automatically take advantage of new architectures that will
McCool: We have seen large speedups on a wide range of applica- likely arise, such as increases in the number of cores.
tions, including database operations, image and video processing,
financial modeling, pattern matching and analysis, many different LJ: Can you tell us more about your recently demonstrated
kinds of scientific computation—the list goes on and on. The support for Intel and AMD multicore CPUs?
RapidMind platform supports a general-purpose programming model DePaul: It’s difficult to overstate the value we bring to software
and can be applied to any kind of computation. The HP test was companies targeting Intel and AMD multicore CPUs. For example,
compute-bound, and it could take advantage of the high compute at SIGGRAPH in San Diego, we demonstrated a 10x performance
performance of GPUs. However, in memory-bound applications, we improvement on an application running on eight CPU cores.
have also seen a significant benefit, over an order of magnitude, RapidMind-enabled applications will scale to any number of cores,
from running the application on RapidMind. RapidMind not only even across multiple processors, and will be tuned for both Intel
manages parallel execution, it also manages data flow and so can and AMD architectures. Software organizations can now target
also directly address the memory bottleneck. As a software platform multicore CPUs, as well as accelerators, such as ATI and NVIDIA
company, we are constantly surprised by the variety of applications GPUs and the Cell processor, all with the same source code.
that developers are RapidMind-enabling. Prior to the launch of our
v2.0 product in May 2007, we had more than 1,000 developers LJ: Is there anything else you’d like to tell our readers?
from many different industries in our Beta program. The problem DePaul: It’s becoming clear that software organizations’ plans for
is industry-wide, and we have developed a platform that has very multicore processors and accelerators will be one of the most
broad applicability. important initiatives they take this year. Companies that choose to
do nothing will quickly find themselves behind the performance
LJ: Shouldn’t this kind of adaptation to multiple cores take curve. Companies that embark on large complex multithreading
place in something more fundamental like the GNU C Library? projects will be frustrated with the costs and timelines, and in the
Is it only a matter of time before such libraries catch up? end, largely disappointed with the outcome. We are fortunate to
McCool: Simply parallelizing the standard library functions would not be partnering with a group of software organizations that see an
have the same benefit, because they do not, individually, do enough opportunity to deliver substantial performance improvements to
work. RapidMind programs, in contrast, can do an arbitrary amount of their customers without a devastating impact on their software
user-specified parallel computation. development cycles.
Although RapidMind looks like a library to the developer,
it’s important to realize that most of the work is done by the LJ: Thank you so much for your time! I
runtime platform. The challenge facing multicore developers is
not one that can be solved solely with libraries. Developers need a Nicholas Petreley is Editor in Chief of Linux Journal and a former programmer, teacher, analyst
system that efficiently takes care of the complexities of multicore: and consultant who has been working with and writing about Linux for more than ten years.

66 | november 2007 w w w. l i n u x j o u r n a l . c o m
 INDEPTH

High-Performance Network
Programming in C
Programming techniques to get the best performance from your TCP applications.
GIRISH VENKATACHALAM

TCP/IP network programming in C on Linux is good fun. All the ensure that you get all the attention at the expense of the other nodes
advanced features of the stack are at your disposal, and you can in the LAN. This is the secret of download accelerators. They open
do lot of interesting things in user space without getting into multiple TCP connections to FTP and HTTP servers and download a file
kernel programming. in pieces and reassemble it at multiple offsets. This is not “playing”
Performance enhancement is as much an art as it is a science. It is nicely though.
an iterative process, akin to an artist gingerly stroking a painting with We want to be well-behaved citizens, which is where non-blocking
a fine brush, looking at the work from multiple angles at different I/O comes in. The traditional approach of blocking reads and writes on
distances until satisfied with the result. the network is very easy to program, but if you are interested in filling
The analogy to this artistic touch is the rich set of tools that Linux the pipe available to you by pumping packets, you must use non-
provides in order to measure network throughput and performance. blocking TCP sockets. Listing 1 shows a simple code fragment using
Based on this, programmers tweak certain parameters or sometimes non-blocking sockets for network read and write.
even re-engineer their solutions to achieve the expected results. Note that you should use fcntl(2) instead of setsockopt(2) for
I won’t dwell further upon the artistic side of high-performance setting the socket file descriptor to non-blocking mode. Use poll(2)
programming. In this article, I focus on certain generic mechanisms or select(2) to figure out when the socket is ready to read or write.
that are guaranteed to provide a noticeable improvement. Based select(2) cannot figure out when the socket is ready to write, so
on this, you should be able to make the final touch with the help watch out for this.
of the right tools. How does non-blocking I/O provide better throughput? The OS
I deal mostly with TCP, because the kernel does the bandwidth schedules the user process differently in the case of blocking and
management and flow control for us. Of course, we no longer have to non-blocking I/O. When you block, the process “sleeps”, which
worry about reliability either. If you are interested in performance and
high-volume traffic, you will arrive at TCP anyway.

What Is Bandwidth?
Once we answer that question, we can ask ourselves another useful
question, “How can we get the best out of the available bandwidth?”
Bandwidth, as defined by Wikipedia, is the difference between the
higher and lower cutoff frequencies of a communication channel.
Cutoff frequencies are determined by basic laws of physics—nothing
much we can do there.
But, there is a lot we can do elsewhere. According to Claude
Shannon, the practically achievable bandwidth is determined by the
level of noise in the channel, the data encoding used and so on.
Taking a cue from Shannon’s idea, we should “encode” our data in
such a way that the protocol overhead is minimal and most of the bits
are used to carry useful payload data.
TCP/IP packets work in a packet-switched environment. We
have to contend with other nodes on the network. There is no
concept of dedicated bandwidth in the LAN environment where
your product is most likely to reside. This is something we can
control with a bit of programming.

Non-Blocking TCP
Here’s one way to maximize throughput if the bottleneck is your local
LAN (this might also be the case in certain crowded ADSL deploy-
ments). Simply use multiple TCP connections. That way, you can Figure 1. Possibilities in Non-Blocking Write with Scatter/Gather I/O

68 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Listing 1. nonblock.c

/* set socket non blocking */ size_t pos = 0;

fl = fcntl(accsock, F_GETFL); ssize_t res;
fcntl(accsock, F_SETFL, fl | O_NONBLOCK);
while (n > pos) {
res = read (sock, buf + pos, n - pos);
void switch ((int)res) {
poll_wait(int fd, int events) case -1:
{ if (errno == EINTR || errno == EAGAIN)
int n; continue;
struct pollfd pollfds[1]; return 0;
memset((char *) &pollfds, 0, sizeof(pollfds)); case 0:
errno = EPIPE;
pollfds[0].fd = fd; return pos;
pollfds[0].events = events; default:
pos += (size_t)res;
n = poll(pollfds, 1, -1); }
if (n < 0) { }
perror("poll()"); return (pos);
errx(1, "Poll failed"); }
}
}
size_t
size_t writenw(int fd, char *buf, size_t n)
readmore(int sock, char *buf, size_t n) { {
size_t pos = 0;
fd_set rfds; ssize_t res;
int ret, bytes; while (n > pos) {
poll_wait(fd, POLLOUT | POLLERR);
res = write (fd, buf + pos, n - pos);
switch ((int)res) {
poll_wait(sock,POLLERR | POLLIN ); case -1:
bytes = readall(sock, buf, n); if (errno == EINTR || errno == EAGAIN)
continue;
if (0 == bytes) { return 0;
perror("Connection closed"); case 0:
errx(1, "Readmore Connection closure"); errno = EPIPE;
/* NOT REACHED */ return pos;
} default:
pos += (size_t)res;
return bytes; }
} }
return (pos);
size_t
readall(int sock, char *buf, size_t n) { }

leads to a context switch. When you use non-blocking sockets, this is what makes it so interesting.
problem is avoided. You can transfer large chunks of data split between multiple
sources/destinations from/to the network. This could be a useful
Scatter/Gather I/O technique, depending upon your application. Listing 2 shows a
The other interesting technique is scatter/gather I/O or using readv(2) code snippet to illustrate its use.
and writev(2) for network and/or disk I/O. When you combine scatter/gather I/O with non-blocking sockets,
Instead of using buffers as the unit of data transfer, an array of things get a little complex, as shown in Figure 1. The code for tackling
buffers is used instead. Each buffer can be a different length, and this this hairy issue is shown in Listing 3.

w w w. l i n u x j o u r n a l . c o m november 2007 | 69
 INDEPTH

A partial write of any buffer can occur, or you can get any combi- We also need to be able to avoid redundant copies and other
nation of a few full writes and few partial writes. Therefore, the while time-consuming CPU operations to squeeze the maximum bandwidth
loop has to take care of all such possible combinations. from the network. A very effective tool for achieving that is the versa-
tile mmap(2) system call. This is a very useful technique for avoiding
mmap(2) Disk I/O the copy-to-buffer cache and, hence, improves performance for net-
Network programming is not all about sockets, however. We still work I/O. But, if you use mmap(2) with NFS, you are asking for trou-
haven’t solved the problem of having to use hard disks, which are ble. Listing 4 shows a code snippet that illustrates the use of
mechanical devices and consequently are much slower than main mmap(2) for both reading and writing files.
memory and even the network in many, if not most, cases (especially
high-performance computing environments). Socket Options and sendfile(2)
You can use some other form of persistent storage, but today, TCP sockets under Linux come with a rich set of options with
none matches the huge storage capacity that hard disks offer. which you can manipulate the functioning of the OS TCP/IP stack.
Currently, most applications on the Internet push several gigabytes A few options are important for performance, such as the TCP
of data, and you end up with heavy storage needs anyway. send and receive buffer sizes:
To test disk performance, type this:
sndsize = 16384;
$ hdparm -rT /dev/sda (/dev/hda if IDE) setsockopt(socket, SOL_SOCKET, SO_SNDBUF, (char *)&sndsize,
(int)sizeof(sndsize));
Check whether you are getting good throughput. If not, enable rcvsize = 16384;
DMA and other safe options using this command: setsockopt(socket, SOL_SOCKET, SO_RCVBUF, (char *)&rcvsize,
(int)sizeof(rcvsize));
$ hdparm -d 1 -A 1 -m 16 -u 1 -a 64 /dev/sda
I am using conservative values here. Obviously, it should be
much higher for Gigabit networks. These values are
determined by the bandwidth delay product.
Listing 2. uio.c Interestingly, I have never found this to be an issue,
so I doubt if this would give you a performance
#include <sys/types.h> boost. It still is worth mentioning, because the TCP
#include <sys/uio.h> window size alone can give you optimal throughput.
#include <unistd.h> Other options can be set using the /proc pseudo-
filesystem under Linux (including the above two), and
size_t unless your Linux distribution turns off certain options,
writeuio(int fd, struct iovec *iov, int cnt) you won’t have to tweak them.
{ It is also a good idea to enable PMTU (Path
size_t pos = 0; Maximum Transmission Unit) discovery to avoid
ssize_t res; IP fragmentation. IP fragmentation can affect not
n = iov[0].iov_cnt; just performance, but surely it’s more important
while (n > pos) { regarding performance than anything else. To avoid
poll_wait(fd, POLLOUT | POLLERR); fragmentation at any cost, several HTTP servers use
res = writev (fd, iov[0].iov_base + pos, n - pos); conservative packet sizes. Doing so is not a very
switch ((int)res) { good thing, as there is a corresponding increase in
case -1: protocol overhead. More packets mean more headers
if (errno == EINTR || errno == EAGAIN) and wasted bandwidth.
continue; Instead of using write(2) or send(2) for transfer, you
return 0; could use the sendfile(2) system call. This provides sub-
case 0: stantial savings in avoiding redundant copies, as bits are
errno = EPIPE; passed between the file descriptor and socket descriptor
return pos; directly. Be aware that this approach is not portable
default: across UNIX.
pos += (size_t)res;
} Advanced Techniques in Application
} Design
return (pos); Applications should be well designed to take full advan-
tage of network resources. First and foremost, using
} multiple short-lived TCP connections between the same
two endpoints for sequential processing is wrong. It will
work, but it will hurt performance and cause several

70 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Listing 3. nonblockuio.c

writeiovall(int fd, struct iov *iov, int nvec) {

int i, bytes;

i = 0;
while (i < nvec) {
do
{
rv = writev(fd, &vec[i], nvec - i);
} while (rv == -1 &&
(errno == EINTR || errno == EAGAIN));

if (rv == -1) {
if (errno != EINTR && errno != EAGAIN) {
perror("write");
}
return -1;
}
bytes += rv;
/* recalculate vec to deal with partial writes */
while (rv > 0) {
if (rv < vec[i].iov_len) {
vec[i].iov_base = (char *)
vec[i].iov_base + rv;
vec[i].iov_len -= rv;
rv = 0;
}
else {
rv -= vec[i].iov_len;
++i;
}
}
}

/* We should get here only after we write out everything */

return 0;

other headaches as well. Most notably, the TCP TIME_WAIT state has
a timeout of twice the maximum segment lifetime. Because the
round-trip time varies widely in busy networks and networks with
high latency, oftentimes this value will be inaccurate. There are other
problems too, but if you design your application well, with proper
protocol headers and PDU boundaries, there never should be a need
to use different TCP connections.
Take the case of SSH, for instance. How many different TCP
streams are multiplexed with just one connection? Take a cue from it.
You don’t have to work in lockstep between the client and the
server. Simply because the protocols and algorithms are visualized
in a fixed sequence does not imply that the implementation should
follow suit.
 INDEPTH

Listing 4. mmap.c

/****************************************** /******************************************
* mmap(2) file write * * mmap(2) file read *
* * * *
*****************************************/ *****************************************/
caddr_t *mm = NULL; fd = open(filename, O_RDONLY, 0);
if ( -1 == fd)
fd = open (filename, O_RDWR | O_TRUNC | O_CREAT, 0644); errx(1, " File read err");
/* NOT REACHED */
if(-1 == fd)
errx(1, "File write"); fstat(fd, &statbf);
/* NOT REACHED */ filelen = statbf.st_size;

/* If you don't do this, mmapping will never mm = mmap(NULL, filelen, PROT_READ, MAP_SHARED, fd, 0);
* work for writing to files
* If you don't know file size in advance as is if (NULL == mm)
* often the case with data streaming from the errx(1, "mmap() error");
* network, you can use a large value here. Once you /* NOT REACHED */
* write out the whole file, you can shrink it
* to the correct size by calling ftruncate /* Now onward you can straightaway
* again * do a memory copy of the mm pointer as it
*/ * will dish out file data to you
ret = ftruncate(ctx->fd,filelen); */

mm = mmap(NULL, header->filelen, PROT_READ | PROT_WRITE,

MAP_SHARED, ctx->fd, 0); bufptr = mm + off;
if (NULL == mm) /* You can straightaway copy mmapped memory into the
errx(1, "mmap() problem"); network buffer for sending */
memcpy(mm + off, buf, len); memcpy(pkt.buf + filenameoff, bufptr, bytes);
off += len;
/* Please don't forget to free mmap(2)ed memory! */ /* Please don't forget to free mmap(2)ed memory! */
munmap(mm, filelen); munmap(mm, filelen);
close(fd); close(fd);

You can make excellent use of available bandwidth by doing things

in parallel—by not waiting for processing to complete before reading
the next packet off the network. Figure 2 illustrates what I mean.
Pipelining is a powerful technique employed in CPUs to speed up
the FETCH-DECODE-EXECUTE cycle. Here, we use the same technique
for network processing.
Obviously, your wire protocol should have the least overhead and
should work without relying much on future input. By keeping the state
machine fairly self-contained and isolated, you can process efficiently.
Avoiding redundant protocol headers or fields that are mostly
empty or unused can save you precious bandwidth for carrying real
data payloads. Header fields should be aligned at 32-bit boundaries
and so should the C structures that represent them.
If your application already is in production and you want to
enhance its performance, try some of the above techniques. It shouldn’t
be too much trouble to attack the problem of re-engineering an
application if you take it one step at a time. And remember, never
trust any theory—not even this article. Test everything for yourself. If
Figure 2. Pipelining your testing does not report improved performance, don’t do it. Also,

72 | november 2007 w w w. l i n u x j o u r n a l . c o m
make sure your test cases take care of LAN, WAN and, if necessary,
satellite and wireless environments.

A Few Words on TCP

TCP has been a field of intense research for decades. It’s an extremely
complex protocol with a heavy responsibility on the Internet. We often
forget that TCP is what holds the Internet together without collapse
due to congestion. IP connects networks together, but TCP ensures
that routers are not overloaded and that packets do not get lost.
Consequently, the impact of TCP on performance is higher than
any other protocol today. It is no wonder that top-notch researchers
have written several papers on the topic.
The Internet is anything but homogeneous. There is every possible
physical layer of technology on which TCP/IP works today. But, TCP is
not designed for working well through wireless networks. Even a high-
latency satellite link questions some of TCP’s assumptions on window
size and round-trip time measurement.
And, TCP is not without its share of defects. The congestion control
algorithms, such as slow start, congestion avoidance, fast retransmit,
fast recovery and so on, sometimes fail. When this happens, it hurts
your performance. Normally, three duplicate ACK packets are sufficient
for triggering congestion control mechanisms. No matter what you do,
these mechanisms can drastically decrease performance, especially if
you have a very high-speed network.
But, all else being equal, the above techniques are few of the most
useful methods for achieving good performance for your applications.

Conclusion
Gunning for very high performance is not something to be taken lightly.
It’s dependent on heuristics and empirical data as well as proven
techniques. As I mentioned previously, it is an art best perfected by
practice, and it’s also an iterative process. However, once you get a
feel for how things work, it will be smooth sailing. The moment you
build a stable base for a fast client/server interaction like this, building
powerful P2P frameworks on top is no great hassle.I

Girish Venkatachalam is an open-source hacker deeply interested in UNIX. In his free time, he likes
to cook vegetarian dishes and actually eat them. He can be contacted at [email protected].

Resources
Polipo User Manual:
www.pps.jussieu.fr/~jch/software/polipo/manual

TCP Tuning and Network Troubleshooting: www.onlamp.com/pub/

a/onlamp/2005/11/17/tcp_tuning.html

Wikipedia’s Definition of Bandwidth:

en.wikipedia.org/wiki/Bandwidth

Advanced Networking Techniques: beej.us/guide/bgnet/output/

html/multipage/advanced.html

TCP and Congestion Control Slides:

www.nishida.org/soi1/mgp00001.html
 INDEPTH

Multiple Associations
with Stream Control
Transmission Protocol
The elegant way SCTP handles multiple streams makes it ideal for things like chat clients.
JAN NEWMARCH

In two previous articles [in the September and October 2007 issues tion then will need to examine each read to see whether it is an event
of LJ], I looked at the basics of SCTP, how you can use SCTP as a or data to be processed. This is quite straightforward. If the flags field
replacement for TCP and how you can use SCTP to process multiple in the sctp_recvmsg() call has the MSG_NOTIFICATION bit set, the read
streams within a single association. In this final article, I look at how message contains an event; otherwise, it contains data as before.
a single endpoint deals with multiple associations (connections to Pseudo-code for this is:
other endpoints). First though, I explain how SCTP can give extra
information about what is going on through events. nread = sctp_rcvmsg(..., msg, ..., &flags);
if (flags & MSG_NOTIFICATION)
Events handle_event(msg);
The SCTP stack can generate events when “interesting” things hap- else
pen. By default, all event generation is turned off except for data handle_data(msg, nread);
events. In the last article, I discussed the SCTP call sctp_rcvmsg(). By
default, this just returns the data read. But, I also wanted to find out Events can be used to tell the following: if a new association has
on which stream the data came, and for this I had to turn on the started or if an old one has terminated; if a peer has changed state by,
data_io_event so the SCTP stack would fill in the sctp_sndrcvinfo say, one of the interfaces becoming unavailable or a new interface
structure, which has the sinfo_stream field. Events are listed in the becoming available; if a send has failed, a remote error has occurred
sctp_event_subscribe structure: or a remote peer has shut down; if partial delivery has failed; and if
authentication information is available.
struct sctp_event_subscribe { If an event is received in the event buffer, first its type must be
uint8_t sctp_data_io_event; found, and then the buffer can be cast to a suitable type for that
uint8_t sctp_association_event; event. For example, the code to handle a shutdown event is:
uint8_t sctp_address_event;
uint8_t sctp_send_failure_event; void handle_event(void *buf) {
uint8_t sctp_peer_error_event; union sctp_notification *notification;
uint8_t sctp_shutdown_event; struct sn_header *head;
uint8_t sctp_partial_delivery_event;
uint8_t sctp_adaptation_layer_event; notification = buf;
uint8_t sctp_authentication_event; switch(notification->sn_header.sn_type) {
}; case SCTP_SHUTDOWN_EVENT: {
struct sctp_shutdown_event *shut;
An application sets fields to one for events it is interested in shut = (struct sctp_shutdown_event *) buf;
and zero for the others. It then makes a call to setsockopt() with printf("Shutdown on assoc id %d\n",
SCTP_EVENTS. For example: shut->sse_assoc_id);
break;
memset(&event, 0, sizeof(event)); }
event.sctp_data_io_event = 1; default:
event.sctp_association_event = 1; printf("Unhandled event type %d\n",
setsockopt(fd, IPPROTO_SCTP, SCTP_EVENTS, notification->sn_header.sn_type);
&event, sizeof(event)); }
}
Events are delivered inline along with “ordinary” data whenever a
read (using sctp_recvmsg or similar) is done. If the application turns on Closing an Association
events, reads will contain a mixture of events and data. The applica- A socket can support multiple associations. If you close a socket, it

74 | november 2007 w w w. l i n u x j o u r n a l . c o m
closes all of the associations! It is sometimes desirable to close only a offers another variation, roughly halfway between TCP and UDP.
single association but not the socket, so that the socket can continue An SCTP socket can handle multiple long-lived associations to
to be used for the other associations. many endpoints simultaneously. It supports the “connection-oriented”
SCTP can abort an association or close it gracefully. Graceful shut- semantics of TCP by maintaining an association ID for each associa-
down will ensure that any queued messages are delivered properly tion. On the other hand, it is like UDP in that each read usually returns
before shutdown, while abort does not do this. Either of these are a complete message from a client. SCTP applications use the TCP
signaled by setting the sinfo_flags in the sctp_sndrcvinfo structure to model by using the one-to-one sockets that I have discussed in the
the appropriate value. A graceful shutdown is signaled by setting the previous two articles. And, it uses a one-to-many model, which is
shutdown flag and writing a message (with no data): more like UDP by using a one-to-many socket. When you create a
socket, you specify whether it is one-to-one or one-to-many. In the
sinfo.sinfo_flags = SCTP_EOF; sctp_send(..., &sinfo, ...); first article in this series, I created a one-to-one socket by the call:

The reader then will be sent an sctp_shutdown_event if it has sockfd = socket(AF_INET, SOCK_STREAM, IPPROTO_SCTP)
that event type enabled. The code to handle such an event was
shown above. This can be done only on one-to-many sockets though. To create a one-to-many socket, I simply change the second
For one-to-one sockets, you are limited to using close(). parameter:

Getting the Association ID sockfd = socket(AF_INET, SOCK_SEQPACKET, IPPROTO_SCTP)

Many of the calls that deal with associations take an association ID as
a parameter. Whereas in TCP, a connection effectively is represented A TCP server handles multiple connections simultaneously by
by the pair of source and destination endpoint IP addresses, in SCTP, essentially using concurrent reads. This is done by using multiple pro-
the source and destination can both be multihomed, so they will be cesses, threads, or by poll/select among many sockets. A UDP server
represented by the set of source and the set of destination addresses. typically uses a single read loop, handling each message as it arrives.
For one-to-many sockets, the source addresses may be shared by An SCTP one-to-many server looks like a UDP server: it will bind a
many associations, so I need the destination addresses to identify an socket and listen. Then, instead of blocking on accept(), which
association properly. For a single association, these destination would return a new one-to-one socket, it blocks on sctp_rcvmsg(),
addresses all belong to a single endpoint computer. The SCTP variation which returns a message from either a new or existing association.
on getsockopt()—that is, sctp_opt_info()—is used to find an associa- Pseudo-code for such a server is:
tion from an address. The reason I cannot simply use getsockopt() is
that I need to pass in a socket address, and the return value includes sockfd = socket(...);
the association value. This in/out semantics is not supported by all bind(sockfd, ...);
implementations of getsockopt(). The code is: listen(sockfd, ...);
while (true) {
sctp_assoc_t get_associd(int sockfd, struct sockaddr *sa, socklen_t salen) { nread = sctp_rcvmsg(sockfd, ..., buf, ..., &info);
struct sctp_paddrinfo sp; assoc_id = sinfo.sinfo_assoc_id;
int sz; stream = sinfo.sinfo_stream;
handle_message(assoc_id, stream, buf, nread);
sz = sizeof(struct sctp_paddrinfo); }
bzero(&sp, sz);
memcpy(&sp.spinfo_address, sa, salen); A client also can use the one-to-many socket model. After binding
if (sctp_opt_info(sockfd, 0, SCTP_GET_PEER_ADDR_INFO, &sp, &sz) == -1) to a port (probably an ephemeral one), it can use the single socket to
perror("get assoc"); connect to many other endpoints and use this single socket to send
return (sp.spinfo_assoc_id); messages to any of them. It even can do away with an explicit connect
} operation and just start sending to new endpoints (an implicit connection
is done if no existing association exists).
Note that Unix Network Programming (volume 1, 3rd ed.) by W.
Richard Stevens, et al., gives different code: the specification has Peeled-Off Sockets
changed since that book was written, and the above is now the pre- One-to-one sockets follow the TCP model; one-to-many sockets follow
ferred way (and Stevens’ code doesn’t work under Linux anyway). the UDP model. Is it possible to have both at once? Yes, it is, to some
extent. For example, you may have a server that you can talk to in
Multiple Associations two modes: ordinary user and superuser. Messages from ordinary
A server can handle multiple clients in a number of ways: a TCP server users may be handled in UDP style, reading and just responding,
can use a single server socket that listens for clients and deals with while superuser connections may need to be treated differently.
them sequentially, or it could fork off each new client connection as a SCTP allows a connection on a one-to-many socket to be “peeled
separate process or thread, or it could have many sockets and poll or off” and become a one-to-one socket. This one-to-one socket may
select between them. A UDP server typically will keep no client state then be treated in TCP-style, while all other associations remain on
and will treat each message in its entirety as a separate entity. SCTP the one-to-many socket.

w w w. l i n u x j o u r n a l . c o m november 2007 | 75
 INDEPTH

Listing 1. chat_client.c

#include <stdio.h> (struct sockaddr *) &serv_addr,

#include <stdlib.h> sizeof(serv_addr)) < 0) {
#include <string.h> perror("connect to server failed");
#include <sys/select.h> exit(3);
#include <sys/types.h> }
#include <sys/socket.h> printf("Connected\n");
#include <netinet/in.h>
while (1) {
#include <netinet/sctp.h> /* we need to select between messages FROM the user
on the console and messages TO the user from the
#define SIZE 1024 socket
char buf[SIZE]; */
#define STDIN 0 FD_CLR(sockfd, &readfds);
char *msg = "hello\n"; FD_SET(sockfd, &readfds);
#define ECHO_PORT 2013 FD_SET(STDIN, &readfds);
printf("Selecting\n");
int main(int argc, char *argv[]) { select(sockfd+1, &readfds, NULL, NULL, NULL);
int sockfd;
int nread, nsent; if (FD_ISSET(STDIN, &readfds)) {
int flags, len; printf("reading from stdin\n");
struct sockaddr_in serv_addr; nread = read(0, buf, SIZE);
struct sctp_sndrcvinfo sinfo; if (nread <= 0 )
fd_set readfds; break;
sendto(sockfd, buf, nread, 0,
if (argc != 2) { (struct sockaddr *) &serv_addr,
fprintf(stderr, "usage: %s IPaddr\n", argv[0]); sizeof(serv_addr));
exit(1); } else if (FD_ISSET(sockfd, &readfds)) {
} printf("Reading from socket\n");
/* create endpoint using SCTP */ len = sizeof(serv_addr);
sockfd = socket(AF_INET, SOCK_SEQPACKET, nread = sctp_recvmsg(sockfd, buf, SIZE,
IPPROTO_SCTP); (struct sockaddr *) &serv_addr,
if (sockfd < 0) { &len,
perror("socket creation failed"); &sinfo, &flags);
exit(2); } write(1, buf, nread);
/* connect to server */ }
serv_addr.sin_family = AF_INET; }
serv_addr.sin_addr.s_addr = inet_addr(argv[1]); close(sockfd);
serv_addr.sin_port = htons(ECHO_PORT); exit(0);
}
if (connect(sockfd,

Lazy Person’s Chat SCTP is a better choice: it can sit in a read loop too, but it also
In this section, I discuss a simple example of how to build a simple chat keeps a list of associations and, better, keeps that list up to date by
server using SCTP. This isn’t meant to be a competitor to the many chat sending heartbeat messages to the peers. The list management is
systems around, rather it is to show some of the features of SCTP. handled by SCTP.
A chat server must listen for messages coming from a probably TCP also could be a choice: each client would start a new client socket
transient group of clients. When a message is received from any one on the server. The server then would need to keep a list of the client sock-
client, it should send the message back out to all of the other clients. ets and do a poll/select between them to see if anyone is sending a mes-
UDP could be a choice here: a server simply can wait in a read sage. Again, SCTP is a better choice: in the one-to-many mode, it will
loop, waiting for messages to come in. But, to send them back out, it keep only a single socket, and there is no need for a poll/select loop.
needs to keep a list of clients, and this is a bit more difficult. Clients When it comes to sending messages back to all the connected
will come and go, so some sort of “liveness” test is needed to keep clients, SCTP makes it even easier—the flag SCTP_SENDALL that can
the list up to date. can be set in the sctp_sndrcvinfo field of sctp_send(). So a server simply

76 | november 2007 w w w. l i n u x j o u r n a l . c o m
needs to read a message from any client, set the SCTP_SENDALL bit
and write it back out. The SCTP stack then will send it to all live peers!
There are only a few lines of code:

nread = sctp_recvmsg(sockfd, buf, SIZE,

(struct sockaddr *) &client_addr,
&len, &sinfo, &flags);
bzero(&sinfo, sizeof(sinfo));
sinfo.sinfo_flags |= SCTP_SENDALL;
sctp_send(sockfd, buf, nread, &sinfo, 0);

The SCTP_SENDALL flag has been introduced only recently into

SCTP and is not in current kernels (up to 2.6.21.1), but it should make
it into the 2.6.22 kernels. The full code for client and server is shown
in Listings 1 (chat_client.c) and 2 (chat_server.c).

Unordered Messages
SCTP normally delivers messages within a stream in the order in which
they were written. If you don’t need this, you can turn off the order-
ing feature. This can make delivery of messages faster, as they don’t
have to be reassembled into the correct order.

New Protocols
I have examined in these three articles how TCP applications can be moved
to SCTP and discussed the new features of SCTP. So, why isn’t everyone
using SCTP now? Well, there is the inertia of moving people off the TCP
applications onto the SCTP versions, and that will happen only when peo-
ple become fed up with the TCP versions—and that may never happen.
The place to look for SCTP is in new applications using new proto-
cols designed to take advantage of SCTP:

I SS7 (Signaling System 7, see Wikipedia) is a standard for control

signaling in the PSTN (Public Switched Telephone Network). SS7
signaling is done out of band, meaning that SS7 signaling mes-
sages are transported over a separate data connection. This repre-
sents a significant security improvement over earlier systems that
used in-band signaling. SCTP basically was invented to handle pro-
tocols like SS7 over IP. SS7 uses multihoming to increase reliability
and streams to avoid the TCP problem of head-of-line blocking.

I Diameter (RFC 3588, www.rfc-editor.org/rfc/rfc3588.txt) is an IETF

protocol to supply an Authentication, Authorization and Accounting
(AAA) framework for applications, such as network access or IP mobility.
A good introduction is at www.interlinknetworks.com/whitepapers/
Introduction_to_Diameter.pdf. It replaces an earlier protocol, Radius,
that ran over UDP. Diameter uses TCP or SCTP for the added reliability
of these transports. A Diameter server must support both TCP and
SCTP; although at present, clients can choose either. SCTP is the default,
and in the future, clients may be required to support SCTP. SCTP is
preferred, because it can use streams to avoid the head-of-line blocking
problem that exists with TCP.

I DLM (Distributed Lock Manager, sources.redhat.com/cluster/dlm)

is a Red Hat project currently in the kernel. This can use either TCP
or SCTP. SCTP has the advantage of multihome support. Although
TCP presently is the default, SCTP can be used by setting a kernel
build configuration flag.
 INDEPTH

Listing 2. chat_server.c

#include <stdio.h> perror(NULL);

#include <stdlib.h> exit(3); }
#include <strings.h>
#include <sys/types.h> bzero(&events, sizeof(events));
#include <sys/socket.h> events.sctp_data_io_event = 1;
#include <netinet/in.h> if (setsockopt(sockfd, IPPROTO_SCTP,
SCTP_EVENTS, &events, sizeof(events))) {
#include <netinet/sctp.h> perror("set sock opt\n");
}
#define SIZE 1024
char buf[SIZE]; /* specify queue */
#define CHAT_PORT 2013
listen(sockfd, 5);
int main(int argc, char *argv[]) { printf("Listening\n");
int sockfd, client_sockfd;
int nread, nsent, len; for (;;) {
struct sockaddr_in serv_addr, client_addr; len = sizeof(client_addr);
struct sctp_sndrcvinfo sinfo; nread = sctp_recvmsg(sockfd, buf, SIZE,
int flags; (struct sockaddr *) &client_addr,
struct sctp_event_subscribe events; &len,
sctp_assoc_t assoc_id; &sinfo, &flags);

/* create endpoint */ printf("Got a read of %d\n", nread);

sockfd = socket(AF_INET, SOCK_SEQPACKET, write(1, buf, nread);
IPPROTO_SCTP); /* send it back out to all associations */
if (sockfd < 0) {
perror(NULL); bzero(&sinfo, sizeof(sinfo));
exit(2); sinfo.sinfo_flags |= SCTP_SENDALL;
}
/* bind address */ sctp_send(sockfd, buf, nread,
serv_addr.sin_family = AF_INET; // (struct sockaddr *) &client_addr, 1,
serv_addr.sin_addr.s_addr = htonl(INADDR_ANY); &sinfo, 0);
serv_addr.sin_port = htons(CHAT_PORT);
if (bind(sockfd, (struct sockaddr *) &serv_addr, }
sizeof(serv_addr)) < 0) { }

I MPI (Message Passing Interface, www.mpi-forum.org) is a de There also is ongoing work in bringing a security model into SCTP, so
facto standard for communication among the processes modeling that, for example, TLS can be run across SCTP. There also is work
a parallel program on a distributed memory system (according to being done on different language bindings to SCTP, such as a Java lan-
Wikipedia). It does not specify which transport protocol should be guage binding. SCTP will not make TCP and UDP disappear overnight,
used, although TCP has been common in the past. but I hope these articles have shown that it has features that can
make writing many applications easier and more robust.
Humaira Kamal, in his Master’s thesis, investigated using SCTP as a Of course, SCTP is not the only attempt to devise new protocols. For
transport protocol and reported favourable results. He singled out the comparisons to other new procotols see “Survey of Transport Protocols
causes as being the message-based nature of SCTP and the use of other than Standard TCP” at www.ogf.org/Public_Comment_Docs/
streams within an association. These examples show that SCTP is Documents/May-2005/draft-ggf-dtrg-survey-1.pdf. This shows
being used in a variety of real-world situations to gain benefits over that SCTP stacks up very well against possible alternatives, so you
the TCP and UDP transports. might want to consider it for your next networking project!I

Conclusion Jan Newmarch is Honorary Senior Research Fellow at Monash University. He has been using
This series of articles has covered the basics of SCTP. There are many Linux since kernel 0.98. He has written four books and many papers and given courses on
options that can control, in fine detail, the behaviour of an SCTP stack. many technical topics, concentrating on network programming for the last six years. His
Web site is jan.newmarch.name.

78 | november 2007 w w w. l i n u x j o u r n a l . c o m
 INDEPTH

Roman’s Law and Fast

Processing with Multiple
CPU Cores
Some practical methods to exploit multiple cores and find thread synchronization problems.
ROMAN SHAPOSHNIK

Computers are slow as molasses on a cold day, and they’ve your PATH, as follows:
been like that for years. The hardware industry is doing its work,
but computers are not going to get any faster unless the software PATH=/opt/sun/sunstudio12/bin:$PATH
industry follows suit and does something about it. Processing
speed is less about GHz and more about multiple cores these My goal is to explain the kind of practical steps you can take to
days, and software needs to adapt. teach that old serial code of yours a few multicore tricks.
I will be so bold as to postulate my own law of the multicore There are three basic steps to iterate through to add parallelism
trend for hardware: the number of cores on a chip would double gradually to your serial application:
every three years. It remains to be seen whether I’m going to be
as accurate in this prediction as Gordon Moore happened to be, 1. Identify parallelism.
but it looks good so far. With Intel and AMD introducing quad-
core systems and Sun pushing the envelope even further with its 2. Express parallelism.
first hexadeca-core CPU named Rock, the question the software
industry has to ask itself is “Are we ready to make all these execu- 3. Measure and observe.
tion threads do useful work for us?” My strong opinion is that we
are not ready. A new paradigm is yet to be invented, and we And, even though the first two steps sound like the most
don’t really know what it should look like. What we do know, exciting ones, I cannot stress enough the importance of step 3.
however, in the words of Herb Sutter, is “The biggest sea change Parallel programming is difficult and riddled with unexpected
in software development since the OO revolution is knocking at performance gotchas. And, there is no other way of being certain
the door, and its name is Concurrency.” that your code got faster than proving it with numbers. The good
The first layer of software where hardware parallelism is going news is that it isn’t all that difficult. If you happen to develop
to percolate is the kernel of your operating system. And, I know mostly for Intel architectures, you can use code similar to the
of only two kernels that have what it takes to tackle the chal- FFMPEG’s START_TIMER/STOP_TIMER for microbenchmarking:
lenge: Solaris and Linux. If there’s any Mac OS X—or dare I say,
Windows—fanboys out there, I have two phrases for you: “256 START_TIMER
processor systems” and “Completely Fair Scheduler”. Now, having do_interesting_stuff();
a kernel that is capable of efficiently multiplexing a large number STOP_TIMER("do_interesting_stuff_tag")
of processes to a large number of hardware threads is only as
good as your demand in actually having that large number of indi- Additionally, there’s the Sun Studio Performance analyzer for
vidual processes running in parallel. And, as much as it is an ISP or observing macro behaviour. You also can use tools like Intel’s
provisioning person’s dream come true, on my laptop I rarely have VTune or even time(1), but whatever you do, make sure that
more than four really active processes running at the same time. performance regression testing is as much a part of your testing
The real need that I have is for each individual application to be routine as the regular regression testing is. You do regression
able to take advantage of the underlying hardware parallelism. testing, don’t you?
And, that’s how a fundamental hardware concept permeates yet Identifying parallelism in an existing application usually starts
another software layer—an application one. At the end of the day with finding spots with data parallel characteristics, task parallel
we, as userland software developers, have no other choice but to characteristics and figuring out a scheduling model to tie the two
embrace the parallel view of the world fully. Those pesky hard- together. Data parallelism usually can be found in applications
ware people left us no choice whatsoever. working with large sets of global or static data (think audio, video
For the rest of this article, I assume that you have at least a and image processing, gaming engines and rendering software).
dual-core system running Linux kernel 2.6.x and that you have Task parallelism, on the other hand, mostly is appropriate when
Sun Studio Express installed in /opt/sun/sunstudio and added to branch-and-bound computation takes place (think Chess-solvers

80 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Advertiser Index
For advertising information, please contact our sales
when a bunch of tasks are asked to do similar calculations, but if department at 1-713-344-1956 ext. 2 or [email protected].
one finds a solution, there’s no need to wait for others). www.linuxjournal.com/advertising
Once you’ve identified all potential sources of the parallelism in
your application, you have to decide what programming tech-
niques to use for expressing it. For an application written in C or
C++, the most commonly used one happens to be explicit paral-
lelization with POSIX threads. This method has been around for Advertiser Page # Advertiser Page #
decades, and most developers usually have some familiarity with it. ABERDEEN, LLC 41 LISA 07 79
On the other hand, given its inherent complexity and the fact that www.aberdeeninc.com www.usenix.org/events/usenix07
it no longer is the only game in town, I’m going to skip over it.
APPRO HPC SOLUTIONS C2 LOGIC SUPPLY, INC. 6
Let’s look at this sample code, which happens to be a very
appro.com www.logicsupply.com
simple routine for calculating how many prime numbers there are
between 2 and N: ASA COMPUTERS 71, 91 LPI 67

www.asacomputers.com www.lpi.org
1 #include <stdio.h>
2 #include <stdlib.h> AVOCENT CORPORATION 1 MICROWAY, INC. C4, 63

3 #include <math.h> www.avocent.com/remotecontrol www.microway.com

4 #include <omp.h>
CARI.NET 83 MIKRO TIK 45
5
www.cari.net www.routerboard.com
6 /* pflag[v] == 1 if and only if v is a prime number */
7 char *pflag; CORAID, INC. 17 POGO LINUX 27

8 www.coraid.com www.pogolinux.com

9 int is_prime(int v)
COYOTE POINT 3 POLYWELL COMPUTERS, INC. 31
10 {
www.coyotepoint.com www.polywell.com
11 int i;
12 int bound = floor(sqrt(v)) + 1; EMAC, INC. 19 THE PORTLAND GROUP 29

13 www.emacinc.com www.pgroup.com

14 for (i = 2; i < bound; i++) {

EMPERORLINUX 13 QSOL.COM 5
15 if (v % i == 0) {
www.emperorlinux.com www.qsol.com
16 pflag[v] = 0;
17 return 0; /* v is NOT a prime number */ FAIRCOM 39 RACKSPACE MANAGED HOSTING C3

18 } www.faircom.com www.rackspace.com

19 }
GENSTOR SYSTEMS, INC. 77 R CUBED TECHNOLOGIES 15
20 return 1; /* v is a prime number */
www.genstor.com www.rcubedtech.com
21 }
22 HPC SYSTEMS, INC. 7 SERVERS DIRECT 9
23 int main(int argc, char **argv) www.hpcsystems.com www.serversdirect.com
24 {
HURRICANE ELECTRIC 85 SILICON MECHANICS 25, 35
25 int i;
www.he.net www.siliconmechanics.com
26 int total = 0;
27 int N = atoi(argv[1]); INFITECH 20, 21 SUPER MICRO COMPUTER, INC. 33
28 int primes[N]; /* array of prime numbers */ www.infi-tech.com www.supermicro.com
29 pflag=(char*)alloca(N);
30 INTEL 37 TECHNOLOGIC SYSTEMS 14

31 for (i = 2; i < N; i++) { www.intel.com www.embeddedx86.com

32 pflag[i] = 1; /* all numbers are prime until... */

KING STAR COMPUTER 73 UNIWIDE TECHNOLOGIES 11
33 } /* ...proven otherwise */
www.kingstarusa.com www.uniwide.com
34
35 for (i = 2; i < N; i++) { /* let the testing begin! */ LINUX JOURNAL 16, 93, 95 WILEY TECHNOLOGY PUBLISHING 23

36 if ( is_prime(i) ) { www.linuxjournal.com www.wiley.com

37 primes[total] = i;
38 total++;
39 }
40 }
41
42 printf("Number of prime numbers between 2 and %d: %d\n",
43 N, total);

w w w. l i n u x j o u r n a l . c o m november 2007 | 81
 INDEPTH

44 generated executable and report runtime statistics for the function

45 return 0; main. Now we know that whatever we do to the source code, the
46 } result shouldn’t be slower than 3.336 seconds. With that in mind,
let’s try asking the compiler to do its best not only at identifying
Granted, the code is silly (some might even say brain-dead), but parallelism (-xloopinfo), but at expressing it as well (-xautopar):
let’s pretend it is a real-life application. In that case, we certainly would
benefit from as much automation as possible. And, if you think about $ cc -fast -xloopinfo -xautopar prime.c -o prime
it, there’s no tool better suited for helping us than a compiler—after "prime.c", line 14: not parallelized, loop has multiple exits
all, it already takes care of understanding the semantics of the code "prime.c", line 14: not parallelized, loop has multiple exits
in order to perform optimizations. Ideally, what we would need is a (inlined loop)
compiler that talks back to us, helping us understand the source code "prime.c", line 31: PARALLELIZED, and serial version generated
better and make reasonable tweaks based on that information. Here’s "prime.c", line 35: not parallelized, unsafe dependence (total)
how Sun Studio 12 lets you do that:
So, with only two extra command-line options, the compiler was
$ cc -g -fast prime.c -o prime smart enough to parallelize the loop on line 31 (-xautopar) and honest
$ er_src prime enough to explain why two other loops (lines 14 and 35) cannot be
..... parallelized easily (-xloopinfo). That’s pretty impressive, but let’s see
Source loop below has tag L3 whether it got us any speedup at all:
35. for (i = 2; i < N; i++) { /* let the testing begin! */
$ export OMP_NUM_THREADS=4
Function is_prime inlined from source file prime.c into the code $ collect ./prime 2000000
for the following line. 1 loops inlined Creating experiment database test.2.er ...
Loop in function is_prime, line 14 has tag L4 Number of prime numbers between 2 and 2000000: 148933
36. if ( is_prime(i) ) { $ er_print -statistics test.2.er | grep Duration
Duration (sec.): 3.331
Finally! Your compiler actually explains to you in plain human
language, what transformations it applied to the source code to Good. It isn’t slower (although not significantly faster either), but
make it faster (-fast). Not only that, but it also identifies and tags all then again, we didn’t have to do anything with the source code. The
key areas (such as loops) that you later can navigate and inspect for compiler did everything for us (except letting the runtime system use
the parallelization potential. Identifying parallelism just got easier. all the way up to four threads by setting the OMP_NUM_THREADS
But what about expressing parallelism? Would it be completely out environment variable to four). Or did it? What about that loop on
of the question to delegate some of that to the compiler as well? line 35? It doesn’t look any more complicated than the one on line
After all, we are too lazy to use POSIX threads (besides, they are like 31. Seems like the compiler is being overly conservative, and we
the GOTOs of parallel programming, anyway). The good news is need to step in and help it a bit. This time, let’s express parallelism
that with the right compiler it is possible. But, before we go there, with OpenMP.
let’s remember the third step from our “three-step parallelization The formal (and boring) definition of OpenMP states that it is “an
program” and establish a performance baseline: API that supports multiplatform shared-memory parallel programming
in C/C++ and Fortran on all architectures, including UNIX platforms
$ cc -fast prime.c -o prime and Windows NT platforms”. Personally, I’d like to think about
$ collect ./prime 2000000 OpenMP as a method of helping the compiler exploit data parallelism
Creating experiment database test.1.er ... in your application when data dependencies get out of hand. In short,
Number of prime numbers between 2 and 2000000: 148933 OpenMP is something you use when -xautopar complains. Given that,
$ er_print -statistics test.1.er for C and C++, OpenMP is expressed through the #pragmas, it is quite
Execution for entire program safe to add these hints (although making sure that suggested parallel
operations don’t have concurrency problems is still your responsibility).
Start Label: Total As with any #pragma, if the compiler doesn’t understand it, it’ll skip
End Label: Total over it. (At the time of this writing, the following freely available Linux
Start Time (sec.): 0.028 compilers support OpenMP 2.5: Intel Compilers, GCC 4.2 and Sun
End Time (sec.): 3.364 Studio 12.)
Duration (sec.): 3.336 So, how do we use OpenMP to boost the compiler’s confidence in
Total LWP Time (sec.): 3.337 the loop on line 35? Simply add the following pragma to line 34:
Average number of LWPs: 1.000
.................................................... 34 #pragma omp parallel for
35 for (i = 2; i < N; i++) { /* let the testing begin! */
The -fast command-line option instructs the Sun Studio C compiler 36 if ( is_prime(i) ) {
to generate the fastest possible code for the same architecture where
the compilation happens. The last two commands actually run the And, don’t forget to add -xopenmp to the set of command-line options:

82 | november 2007 w w w. l i n u x j o u r n a l . c o m
$ cc -fast -xloopinfo -xautopar -xopenmp prime.c -o prime
"prime.c", line 14: not parallelized, loop has multiple exits
At the end of the day we, as
"prime.c", line 14: not parallelized, loop has multiple exits userland software developers, have
(inlined loop)
"prime.c", line 31: PARALLELIZED, and serial version generated
no other choice but to embrace the
"prime.c", line 35: PARALLELIZED, user pragma used parallel view of the world fully.
Nice! We’ve got two out of three loops in our application now
completely parallelized. Let’s see how much faster it runs: than POSIX threads do, is still not a panacea. After all, even our
simplistic usage of it was enough to introduce a bug. It seems that
$ collect ./prime 2000000 regardless of how we express parallelism, what we need is a tool
Creating experiment database test.3.er ... that would help us uncover concurrency bugs.
Number of prime numbers between 2 and 2000000: 146764 I know of two such tools freely available on Linux: Intel Thread
$ er_print -statistics test.3.er | grep Duration Checker and Sun Studio Thread Analyzer. And, here’s how you can
Duration (sec.): 1.132 use the latter one to combat data races (note that we need an
extra compile-time command-line option -xinstrument=datarace to
No doubt, 294% is an impressive speedup, but how come we make thread analysis possible and that we have to ask collect for
lost some prime numbers? We now have 146,764 of them reported recording data race events by specifying -r on):
instead of 148,933. Maybe the compiler was right in being
conservative and not parallelizing that pesky loop. Should we go $ cc -fast -xloopinfo -xautopar -xopenmp -xinstrument=datarace
back and remove our OpenMP pragma? Not yet. Even though ¯prime.c -o prime
we’ve just caught a bug in our parallelized version of the same $ collect -r on ./prime 2000000
application (which only goes to show how easy it is to introduce Creating experiment database tha.1.er ...
bugs and how much more important
regression testing becomes when you try to
parallelize anything), we still are on the
right track. The question is, how do we
find the actual problem?
The trouble with parallel programming
is that it makes some sections of your
application nondeterministic. That means
now you have to deal with all sorts of
problems you didn’t have to deal with
before, such as race conditions, and dead-
lock and resource allocation issues are the
chief nemeses. The amount of nondeter-
minism introduced is, in fact, something
that makes POSIX threads quite fragile in
most real-life situations—so much so,
that one of the key parallel computing
researchers, Professor Edward A. Lee, made
a pretty scary prediction in his article “The
Problem with Threads”:

I conjecture that most multithreaded

general-purpose applications are, in
fact, so full of concurrency bugs that
as multicore architectures become
commonplace, these bugs will begin to
show up as system failures. This scenario
is bleak for computer vendors: their
next generation of machines will become
widely known as the ones on which many
programs crash.

As you can see, OpenMP, even though it

introduces significantly less nondeterminism

w w w. l i n u x j o u r n a l . c o m november 2007 | 83
 INDEPTH

Number of prime numbers between 2 and 2000000: 148933 $ cc -fast -xloopinfo -xautopar -xopenmp prime.c -o prime
$ er_print -races tha.1.er $ collect ./prime 2000000
No race information recorded in experiments Creating experiment database test.4.er ...
Number of prime numbers between 2 and 2000000: 148933
Weird. Not only did we get the correct result, but also the thread $ er_print -statistics test.4.er | grep Duration
analyzer didn’t seem to notice anything unusual. Is it broken? Not Duration (sec.): 1.130
really. You see, what makes concurrent bugs so notoriously difficult
to track down is the fact that most of them are intermittent. As with It’s 2.95 times faster. Not bad for 15 minutes of work and four
most manifestations of a nondeterministic behavior, they come and go extra lines of OpenMP pragmas giving hints to the compiler!

The trouble with parallel A Few Loose Ends

OpenMP and -xautopar seem to work pretty well for C, but what
programming is that it about C++? Will they mesh well with the kind of modern C++ usage
makes some sections of your peppered with generics and template metaprogramming? The short
answer is, there’s no short answer. But, let’s see for ourselves with the
application nondeterministic. following example of modern C++ [ab]use:

depending on how applications are run, what else runs on the system #include <vector>
and whether you use tools such as a conventional debugger. Thread #include <iterator>
analyzer reports only those problems that did actually occur. Well, if at
first you don’t succeed:
Resources
$ collect -r on ./prime 2000000
Creating experiment database tha.2.er ... Moore’s Law:
Number of prime numbers between 2 and 2000000: 114833 www.intel.com/technology/mooreslaw/index.htm
$ er_print -races tha.2.er
Total Races: 2 Experiment: tha.2.er Sun Studio Express: developers.sun.com/sunstudio/downloads/
express/index.jsp
Race #1, Vaddr: (Multiple Addresses)
Access 1: Write, main -- MP doall from line 34 FFMPEG: ffmpeg.mplayerhq.hu
[_$d1B34.main] + 0x00000172, line 37 in "prime.c"
Access 2: Write, main -- MP doall from line 34 Intel VTune: www.intel.com/cd/software/products/asmo-na/
[_$d1B34.main] + 0x00000172, line 37 in "prime.c" eng/239145.htm
Total Traces: 1
Intel Thread Checker: www.intel.com/cd/software/products/
Race #2, Vaddr: 0xbffff28c asmo-na/eng/291669.htm
Access 1: Write, main -- MP doall from line 34
[_$d1B34.main] + 0x00000189, line 38 in "prime.c" TotalView Debugger: www.totalviewtech.com/index.htm
Access 2: Write, main -- MP doall from line 34
[_$d1B34.main] + 0x00000189, line 38 in "prime.c" POSIX Threads:
Total Traces: 1 www.mhpcc.edu/training/workshop2/pthreads/MAIN.html

Bingo! We reproduced the bug and our tool dutifully reported OpenMP: www.openmp.org
the actual location of where the race condition happened: lines 37
and 38. Things go wrong when two threads find prime numbers and Effective Use of OpenMP in Games:
they try to update the primes array and total variable—a textbook https://www.cmpevents.com/Sessions/GD/EffectiveUse.ppt
example of a race condition. But, it’s pretty easy to fix. We have to
serialize threads entering these two lines of code. Can we do that with Intel Thread Building Blocks: osstbb.intel.com
OpenMP? Sure we can:
Cilk: supertech.csail.mit.edu/cilk
37 #pragma omp critical
38 { “The Problem with Threads”:
39 primes[total] = i; www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-1.pdf
40 total++;
41 } “The Free Lunch Is Over: A Fundamental Turn Toward Concurrency
in Software”: gotw.ca/publications/concurrency-ddj.htm
With that, let’s see what the final speedup is going to be:

84 | november 2007 w w w. l i n u x j o u r n a l . c o m
#include <algorithm> Fibonacci numbers, is something that really makes me sad. Tasks as
#include <iostream> defined in the upcoming OpenMP 3.0 standard seem far less threatening.
void standard_input_sorter() {
using namespace std; Conclusion
vector<string> v; There is a fundamental trend toward concurrency in hardware.
copy(istream_iterator<string>(cin), istream_iterator<string>(), Multicore systems are now making their way into laptops and desk-
back_inserter(v)); tops. Unfortunately, unless software engineers start taking these
sort(v.begin(), v.end()); trends into account, there’s very little that modern hardware can do
} to make individual applications run faster. Of course, parallel pro-
gramming is difficult and error-prone, but with the latest tools and
$ CC -c -fast -xloopinfo -xautopar -xopenmp -library=stlport4 sorter.cc programming techniques, there’s much more to it than merely POSIX
threads. Granted, this article scratches only the surface of what’s
The above produces a pretty long list of complaints, explaining why a available. Hopefully, the information presented here will be enough
particular section of the STLport library cannot be parallelized. The key of a tipping point for most readers to start seriously thinking about
issue here is that certain areas of C++ are notoriously difficult to paral- concurrency in their applications. Our high-definition camcorders
lelize by default. Even with OpenMP, things like concurrent container demand it and so does every gamer on earth.I
access are much more trouble than they are worth. Do we have to
rewrite STL? Well, seems like Intel almost did. Intel has been working on Roman Shaposhnik started his career in compilers back in 1994 when he had to write a translator
what it calls the Thread Building Blocks (TBB) C++ library, and its claim to for the programming language he’d just invented (the language was so weird, nobody else wanted
fame is exactly that—making modern C++ parallel. Give it a try, and see the job). His first UNIX exposure was with Slackware 3.0, and he’s been hooked ever since.
if it works for you. I especially recommend it if you’re interested in Currently, he works for Sun Microsystems in the Developer Products Group. He is usually found
exploiting task parallelism. But, then again, the amount of modern C++ pondering the question of how to make computers faster yet not drive application developers
that TBB throws at even the simplest of examples, such as calculating insane. He runs a blog at blogs.sun.com/rvs and can be reached via e-mail at [email protected].
 INDEPTH

High-Performance
Linux Clusters
The present and future of high-performance computing. DAVID MORTON

Twice a year, a group of scientists in Europe and the United States about understanding technical minutiae related to interconnects,
release a list of the world’s 500 most powerful computing systems. processors and operating systems. Early vendors of high-performance
The Top 500 list is the most prestigious ranking of its kind, with ven- Linux clusters did not necessarily have the expertise in computational
dors and users leveraging favorable rankings to promote their work. fluid dynamics (CFD), finite element analysis (FEA) and visualization
The most recent list, released June 2007, reconfirmed a recent trend: codes of proprietary systems vendors.
Linux is by far the most frequently used operating system in high- It is, therefore, natural for many to question whether the tremen-
performance computing (HPC). Consider the numbers: 389 machines dous price advantage of Linux and open systems still outweighs all other
(or 78%), run some flavor of Linux, 64 run UNIX, two run Windows considerations. The truth is that although Windows provides some
and 42 feature a mix of Linux and other operating systems. advantages to entry-level HPC users, high-performance Linux clusters
Although such dominance suggests that Linux has had a long have matured. Today’s Linux clusters deliver better performance at a
history in HPC, the truth is that Linux clusters began replacing UNIX more attractive price than ever before. Clusters are increasingly being
systems only six years ago. The reason for such quick initial take-up demanded as turnkey systems, allowing faster time to production and
is due to the fact that Linux and open systems brought commodity fewer management headaches. In addition, the very nature of open source
hardware and software into what had previously been a proprietary has contributed to the strength of high-performance Linux clustering.
systems market. This change brought costs down significantly, allow- Linux clusters adapt more quickly to new technology changes, are easier
ing users at the high end to purchase more power at lower cost and to modify and optimize and benefit from a worldwide community of
opening the door for new users, such as traditional product designers developers interested in tweaking and optimizing code.
who were not able to afford closed proprietary systems. The domina-
tion of Linux in the HPC market is so successful that market research The Advantages of Linux-Based HPC
firm IDC estimated Linux represented 65% of the total HPC market by The most important factor in HPC is, of course, performance. National
mid-2006 (as compared to approximately 30% for UNIX), with addi- laboratories and universities want ever-more powerful machines to
tional growth projected. The Top 500 list confirms that growth. solve larger problems with greater fidelity. Aerospace and automotive
engineering companies want better performing systems in order to
Challenges and Questions grow from running component-level jobs (such as analyzing the stress
Linux is clearly the present of HPC, but is it the future? Microsoft contin- on an engine block) to conducting more complex, multi-parameter
ues to make advancements with its Windows Compute Cluster Server, has studies. Product designers in a variety of other fields want to graduate
plenty of cash on hand and is clearly capable, from a business perspective, from running CAE applications on their relatively slow workstations in
of eating up market share. In addition, despite its well-known flaws, order to accelerate the overall design process.
everyone has worked with and is familiar with Windows, potentially Performance, therefore, cannot be separated from high-performance
making it a comfortable platform to new HPC users. computing and in this area, Linux clusters excel. There are two primary
Complicating matters further is that, despite their well-earned reasons for this: maturity and community.
market dominance, high-performance Linux clusters have, in many
cases, earned a reputation for being difficult to build and manage. Maturity
Widely available commodity components lead to complexity in the With years of experience under their belts, vendors and architects of
selection, integration and testing required when building a stable system. high-performance Linux clusters are better equipped than ever to
This complexity becomes doubly problematic when you consider that design stable, tuned systems that deliver the desired price performance
organizations invest in HPC systems in order to get the best possible and enable customers to get the most out of their application licenses.
performance for the applications they run. Small variations in system First-generation systems may have been difficult to manage, but
architecture can have a disproportionately large impact on time to the newest generation comes equipped with advanced cluster man-
production, system throughput and the price/performance ratio. agement software, greatly simplifying operations. By selecting an expe-
Furthermore, like any new technology, the first high-performance rienced vendor, many of today’s clusters are delivered as full-featured
Linux clusters hit bumps in the road. Early systems took a very long systems as opposed to an unwieldy pile of stitched-together commodity
time for vendors to build and deliver and an even longer time to components. As a result, users benefit from both lower acquisition
put into production. Additionally, early management software made costs and easy-to-use high-performance systems.
re-provisioning systems and upgrading components cumbersome. The maturity of the Linux HPC industry also contributes to a deeper
Finally, delivering HPC systems is as much about understanding the understanding of the codes users rely on, as well as the hardware that
nuances of computer-aided engineering (CAE) applications as it is goes into building a system. Certain vendors have become experts at

86 | november 2007 w w w. l i n u x j o u r n a l . c o m
tuning systems and optimizing Linux to meet and overcome the Second, Linux clusters evolved without headless operations. As a
challenges posed by widely used HPC applications. For example, result, administrative tools are able to install and manage the system
most high-performance structures codes, such as those from ANSYS as a whole, rather than as individual workstations or servers. These
or ABAQUS, require high I/O to sustain higher rendering rates. tools continue to get easier to use, enabling users with limited
Conversely, crash/impact codes don’t require much I/O to run optimally; technical skills to jump quickly into HPC. To take just one example,
they are designed to run in parallel in systems where the average CPU Linux Networx recently launched its newest cluster management
count is 16. Linux has evolved to the point where it is now very easy application, Clusterworx Advanced. This application provides system
for vendors to build systems that accommodate the needs of these administrators with intuitive tools that greatly simplify operations and
codes—even within the same cluster. reduce administration workload.
Alliant Techsystems (ATK) is a recent example of how high-performance Third, Linux-based clusters are easy to scale due, in part, to newer
Linux clusters have matured. ATK is an advanced weapon and space filesystems, such as GPFS and Lustre, which provide better scalability,
systems company with many years of experience working with HPC but only on Linux and UNIX. Windows-based filesystems are typically
systems. In 2006, faced with upgrading its aging proprietary system, tuned for file sharing and don’t provide the type of performance and
the launch system’s group invested, after extensive benchmarking, in a accessibility required when lots of compute nodes all request the same
high-performance Linux cluster—finding one tuned and optimized for dataset at the same time.
CFD, FEA and visualization codes. The decision reflected their under- Fourth, resource management tools, such as Altair’s PBS Pro and
standing that Linux clusters—and vendors—had matured. Platform LSF, ensure that computing resources are being allocated with
“We had heard several horror stories of organizations that moved utilization rates exceeding 90%. Without proper resource management,
to Linux supercomputers, only to suffer through installation times systems tend to work only when the engineering team works, thereby
that stretched to six or eight months and beyond”, said Nathan limiting overall utilization. With mature resource management tools, such
Christensen, Engineering Manager at ATK Launch Systems Group. “For as those available for Linux-based HPC systems, jobs can be scheduled 24
instance, one of ATK’s other business units experienced eight weeks of hours a day, 365 days a year. Multiple jobs can be run simultaneously, as
waiting and downtime to get a system into production. The Launch needed, thereby ensuring excess power is always put to use.
Systems Group wanted to avoid a similar experience.” Fifth, from a stability perspective, Linux—due to its flexibility and
“The system arrived application-tuned, validated and ready for the number of people working on refining it—is significantly more
production use”, said Christensen. “We were able to move quickly stable and scalable than other platforms. Windows, for instance, is
into full production, generating our simulations and conducting our prone to failure at moderately larger node counts and is not consid-
analysis within two weeks of delivery.” ered as an option at government and national laboratories.
The system also accelerated the company’s time to results, thereby Sixth, the nature of open source makes Linux the most convenient
enabling ATK to complete designs faster and conduct more frequent, platform for vendors and users to work with. Standards are broadly
higher-fidelity analysis. The launch system’s group completes runs defined and supported by a worldwide community of programmers—
three to four times faster than before. In addition, on some of its key rather than the diminishing numbers found at the remaining propri-
CFD and FEA applications, ATK has been able to achieve ten times the etary vendors. As a result, there are no shortages of fully developed
throughput performance. tools, utilities and software modifications that users and vendors can
leverage in order to optimize their systems.
Community
The greater Linux community is also an important factor in assuring Conclusion
that Linux-based systems deliver the greatest performance. The benefit The HPC market has made its choice, and the choice is the Linux OS
of being open source means that users and vendors from around the due to its superior performance, lower costs and its community of
world continue to develop innovations and share them with the open-source developers and vendors. Windows may have a lot to offer
greater community. This enables Linux-based HPC systems to adapt entry-level users, especially those with more limited resources or goals.
more quickly to new hardware and software technologies. As a result, Likewise, UNIX still has a lot to offer for many legacy HPC applications.
the ability to take advantage of new processors, interconnects and However, both Windows and UNIX require more work to be able to
applications is much greater than with proprietary systems. deliver the same functionality and compelling price-performance value
of Linux. The HPC market is more open and competitive than it has
Additional Benefits ever been, but it is clear that Linux is still the best choice for today and
High-performance Linux clusters offer a range of benefits beyond raw the foreseeable future.I
application performance.
First, Linux is well known for its ability to interoperate with all David Morton brings 17 years’ experience in supercomputing in both vendor and end-user roles to
types of architectures and networks. Because of the investment in HPC Linux Networx. Dave is responsible for leading the Linux Networx technology vision and directing
systems, users want to make certain that their systems are as future- the hardware, software and system engineering teams. Previously, Dave served as technical
proof as possible. Linux provides users with an operating system that is director for the Maui High Performance Computer Center where he was responsible for the definition
flexible enough to accommodate virtually any future advancement. and oversight of technology for this Department of Defense supercomputing center. He held several
This is further amplified, of course, when the larger Linux community, positions at SGI, including director of server I/O and continuation engineering and director of
working together to solve common problems, is again taken into Origin platform engineering. Dave’s experience includes eight years at Cray Research, where
question. In addition, a variety of tools, such as Samba, allow Linux to he was named inventor on three issued patents. Dave earned a Master’s degree in Mechanical
share file services with Windows systems, and vice versa. Engineering from the University of Illinois and an MBA from the University of Minnesota.

w w w. l i n u x j o u r n a l . c o m november 2007 | 87
 INDEPTH

Open-Source Compositing in
Blender
How to use node-based compositing in Blender. DAN SAWYER

Linux is bursting with multimedia poten-

tial—at least, that’s the impression one gets
from the plethora of multimedia-oriented dis-
tributions that have surfaced in recent years.
DeMuDi, Planet CCRMA, Ubuntu Studio,
64 Studio and the list continues on ad infini-
tum. However, for many years now, the term
multimedia proved deceptive. GIMP and
Inkscape and other graphics tools meant that
2-D graphics were covered, and the astound-
ing variety of audio tools available with
real-time priority meant that users’ needs for
recording and audio processing were met.
Video tools lagged behind, because video
processing is both more difficult and more
prone to patent encumbrance. In the last few
years, things have begun to catch up to the
point where it’s feasible to create films or
cartoons from concept through execution
using only Linux tools.
Elephants Dream, one such cartoon, was
the foundation for a major breakthrough in
open-source video. Financed by presales of
a then-unseen cartoon, Elephants Dream Figure 1. Adobe After Effects Interface
was a strategy for raising money to advance
the development of and raise awareness but the prices run up to thousands of dollars which dramatically effect work flow. The
for the open-source 3-D suite Blender per seat with extra costs for maintenance first of these paradigms is the Photoshop
(www.blender.org). In order to accomplish and render-node licenses. For those with paradigm. It shows up most prominently in
this goal, the creators had to develop some- smaller budgets, Linux compositing has After Effects, and in a modified form in
thing that never had been available before: been perpetually out of reach, unless one Apple Motion, and works by marrying the
an open-source compositor. has the patience to do hand scripting in interface conventions of Photoshop with a
Compositing is the art of taking multiple ImageMagick, which is far more technical basic multitrack editor interface. In this
image sources—whether from 3-D, vector than most artists care to get and generally paradigm, composites are achieved using lay-
graphics, photographs, video or procedu- requires the addition of a hacker to the ers of images built atop each other, with
rals—and marrying them together to create a pipeline (another cost point), or work each layer being operated upon by its own
seamless, integrated image. A good com- frame by frame in GIMP, which is laborious effects stack. The main advantages of this
positing program provides the means to and not worth the effort for any but the paradigm are the speed of work for simple
access all the mathematical functions avail- smallest projects. and moderately complex effects and the ease
able in the image processing universe, and a Consequently, the only budget-friendly of navigation for new users who already are
good artist needs to be able to get down solutions for a small studio has been Adobe’s familiar with Photoshop (Figure 1).
into the guts of an image from time to time, After Effects or Apple’s Motion, which means The second paradigm, the “node-based”
below the interface, and tweak it directly adding a Windows or Mac OS machine to paradigm, is the one that appears in the
with mathematical functions. the pipeline. Both are very capable, useful high-end professional compositors. It works
Because of Linux’s continuing adoption in tools that produce professional results, but by chaining together various image functions
post houses, several high-end compositing neither are open source. to create complex effects. Image functions
systems, such as Shake, D2 Nuke and Eyeon The two classes of compositors are built are mathematical transforms applied to an
Fusion, have been available for years now, around two divergent interface paradigms, image to change it in one way or another,

88 | november 2007 w w w. l i n u x j o u r n a l . c o m
Figure 2. A Node-Based Interface

and they reside at the base of anything one

does in GIMP or ImageMagick or in video
compositing. These functions are encapsulat-
ed in the interface by nodes. A node works a
bit like a guitar pedal—it accepts inputs and
generates outputs, and those outputs can be
routed to an infinite number of other nodes.
Thus, in a node-based compositor, one uses
the node chains to accomplish one’s goal,
and there typically are two types of nodes
from which to choose. One type is the famil-
iar, user-friendly prepackaged effects plugins,
such as one would find in the Photoshop
universe. The other type is a set of mathe-
matical interfaces that allow you to build
custom effects yourself. This has the disad-
vantage of being far more visually complex
and, for some people, much harder to learn.
However, for that steeper learning curve, the
artist gets a much more versatile work flow, Figure 3. Finding the Nodes Editor
which is better suited to doing highly com-
plex work. Node-based compositors available it lacks the ability to nest compositions or To demonstrate what it can do, let’s walk
for Linux include: Shake (now defunct), layer effects as complexly as can After Effects through a fairly simple five-layer composite.
Eyeon Fusion, D2 Nuke (formerly of Digital and Motion, leaving it firmly in the quick- To get to the compositor, fire up Blender
Domain, now owned by the Foundry) and and-dirty category for compositing work. and change the main 3-D window to the
Blender (Figure 2). However, with version 2.43, Blender nodes editor (Figure 3). Toggle on the Use
Blender itself has long had a rudimentary introduced its node-based compositor, the Nodes button. Because Blender uses nodes
track-based compositing system, which has jewel in the crown of the Elephants Dream for material and texture creation as well as
received a lot of attention since Elephants improvements. Professional-level open-source compositing, you need to depress the picture
Dream and has become quite useful both as compositing has arrived, and it’s integrated icon. By default, a render scene and a com-
a video editor and a compositor. Alas, with an otherwise very powerful 3-D content posite node will pop up. In the case of this
because its primary purpose is video editing, creation suite. project, one of the elements I’m using is a

w w w. l i n u x j o u r n a l . c o m november 2007 | 89
 INDEPTH

Professional-level
open-source
compositing has
arrived, and it's
integrated with
an otherwise very
powerful 3-D content
creation suite.
Figure 4. Basic Composite Nodes Setup

easier to work with). Now we come to the color keying. There

First, I build the nodes chain for the are a number of ways to do color keying in a
background image (Figure 6). The first node, node-based compositor, not least among
moving left to right, is our source node—the them is building your own keyer out of basic
photo itself. The second is a scale node, mathematical functions. Although this
which I use to shrink my rather large digital method is excellent and yields the best
photograph down to something approximat- results (even on DV footage, which has some
ing SD video resolution. Next, I’ve used an very particular problems with color keying), it
RGB curves node to blow the highlights out requires far more ink than I have space for
of the photo to help sell the illusion that our here. So, for the sake of brevity, I selected
character is looking from a shaded courtyard one of Blender’s three native keying nodes,
out into a wild garden in direct sunlight. the channel key, to do the job (Figure 8).
Next, I take the pillars element, which is The footage, first off, needs some prep,
a rendered layer from within Blender proper, and I’ve prepared the same footage two dif-
add a procedural glow layer to it, and marry ferent ways—one for pulling the matte and
the glow and the pillars to the background. the other for color matching. I first ran the
To do this, I take an output from the source footage through a scaling node to correct for
and run it through a curves pass to cut out the 16:9 aspect ratio I shot in—as the rest of
all but the brightest highlights (Figure 7). my elements are in 4:3, I’m pre-correcting
I pipe the output from the curves node the footage rather than doing it at render
Figure 5. The Viewer Node and Window into a blur node, where I do a 40-pixel x/y time. I then ran it through a translate node,
tent blur and then direct that into a Screen which allowed me to reposition the footage
3-D scene in Blender, so I retain this node and node, where the glow is composited back to the left, so that we actually can look over
assign it to my primary camera (Figure 4). over the source picture of the pillars. This the actress’ shoulder rather than just staring
Next, I split the bottom view into two married image is then piped into an at the back of her head. From there, I send
windows, and in the right-most pane, pull up AlphaOver node, which pastes the pillars and the output into two parallel subtrees—keying
the image editor window, where there is a glow over the top of the photo. and color correction.
list box that allows you to choose the output
nodes from the compositor window. This is
how you check your progress (Figure 5).
Next, I add a few more sources. In each
case, pressing space brings up a menu that
allows me to add nodes. I want three addi-
tional input nodes, and from each I assign
the appropriate type. For this project, I’m
working with two still images (the lens flare
and the background photo) and one image
sequence (the greenscreen clip, rendered out
from DV to an image sequence to make it Figure 6. The Background Nodes Tree

90 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Why an Image
Sequence Instead
of a Video?
Working with image sequences offers three distinct
advantages over working with video. First, with
image sequences it’s easier to retouch problems on
individual frames without reprocessing an entire clip.
Second, the variety of alpha-enabled lossless formats
gives the operator a far greater flexibility than do
video formats, very few of which are lossless or allow
embedded alpha channels. Third, video codecs are,
frankly, a big pain in the neck. They vary wildly in
support from one program to another; however,
image formats are universal and comparatively open.
Converting your video to image sequences before
piping it through your compositor or motion tracker
means that you’re going to encounter far fewer
problems with moving between various programs in
your work flow.

With Blender, there is one further advantage to using

image sequences, and it has to do with a shortcom-
ing of the program. Blender does not support NTSC
drop or nondrop frame encoding, both of which
run at 29.97 frames per second (fps). Because it’s a
European program and has historically been an ani-
mation program, the closest approximation it can hit
is 30fps. Though .3fps seems like a trivial difference,
it’s more than enough to slip your sound sync beyond
the limits of intelligibility, so it’s far better to work
with image sequences and then re-multiplex your
audio with your video in your video editing software,
such as KDENLIVE or OpenMovieEditor.

The keying subtree begins with a curves node, which pushes

the green in the greenscreen into a narrow band to make it easier
for the keyer to latch on to. Then, the output links to the input of
a channel keyer, set to pull the cleanest possible matte (which I
accomplished by hooking a viewer node to the image output, so I
could see what I was doing when I played with the settings). The
resulting matte is then run through a blur node. Normally, when
keying DV footage, I would apply a 4x2 blur to soften the matte
and compensate for the edge artifacting introduced by the DV
compression. However, in this case, my edges were dark, because
of how I lit the original scene, and I needed some extra feathering
so the brightness from the background would bleed over. The
output of this blur node is then piped into the Fac input of an
AlphaOver node, which marries the greenscreen footage to the
rest of the image. But, I’m getting ahead of myself.
Let’s back up to the other half of the keying tree. This takes an
 INDEPTH

Converting your video

to image sequences
before piping it through
your compositor or
motion tracker means
that you're going to
encounter far fewer
problems with
moving between
various programs
in your work flow.
Figure 7. Pillars and Glow Pass
additional output from the translate node
into a curves node, which is set to tamp
down the green channel to get rid of the
green spill and help sell the different lighting
conditions of the foreground vs. the back-
ground. The output of this curves node is
then run into the bottom input on
AlphaOver. Now, to complete the marriage
of foreground with background, we run an
additional noodle from the AlphaOver node
at the end of the background subtree into
the top image input on the keyer AlphaOver
node.
I could leave things here, but the shot
could use a little extra touch to tie all the
layers together. To accomplish this, I created
a nice lens flare and brought it in to Blender. Figure 8. The Color Keying Nodes Tree
I ran it through a translate node to put it into
the right spot, and from there into another
screen node, which lays it over the top of the
previous composite. To do this, the lens flare
went into the top image input, and the pre-
vious AlphaOver node went into the bottom
image input, and I messed with the Fac, so
I got the right effect—just a hint of extra
brightness and anamorphic smear, which
helps sell the integration of the different
layers (Figure 9).
Now, all that remains is to hook up the
Composite node, which is what Blender
draws from for its output. This can be found
next to the viewer node under output in the Figure 9. Layers of the Image Including Lens Flare
add nodes menu, which you get to by
pressing the spacebar. Once the composite click Anim (Figure 10). The result of your directory that you have set for output in
node is hooked up to the output, go to the hard work appears in the render window, the renderbuttons window.
renderbuttons window at the bottom of the from where you can save it using F3 (if it’s Simple though this project is, it gives a
screen, depress the Do Composite button, a still). Or, you can find it on your hard good grounding in how nodes work and why
and click Render, or, if it’s an animation, drive in the temp directory or another they’re useful. Enough access to basic

92 | november 2007 w w w. l i n u x j o u r n a l . c o m
 INDEPTH

Figure 10. Click Render or Anim in the Blender Controls

production to Elephants Dream, gives us more such innovations that

push the compositing system to the next plateau. Until then, there is
much to explore, learn and use.
Open-source compositing has finally arrived. Enjoy!I

Dan Sawyer is the founder of ArtisticWhispers Productions (www.artisticwhispers.com), a small

audio/video studio in the San Francisco Bay Area. He has been an enthusiastic advocate for free and
open-source software since the late 1990s, when he founded the Blenderwars filmmaking community
(www.blenderwars.com). Current projects include the independent SF feature Hunting Kestral and
The Sophia Project, a fine-art photography book centering on strong women in myth.

Statement of Ownership, Management, and Circulation

1. Publication Title: Linux Journal PO Box 980985 10. Owner(s):
2. Publication Number: Houston, TX 77098 Carlie Fairchild
1075-3583 9. Full Names and Complete PO Box 980985
3. Filing Date: September 10, 2007 Addresses of Publisher, Editor, Houston, TX 77098
4. Issue Frequency: Monthly and Managing Editor: Joyce Searls
5. Number of Issues Published Publisher: Carlie Fairchild PO Box 980985
Annually: 12 PO Box 980985 Houston, TX 77098
6. Annual Subscription Price: $25 Houston, TX 77098 Adele Soffa
7. Complete Mailing Address of Editor: Doc Searls PO Box 980985
Figure 11. The Completed Project Known Office PO Box 980985 Houston, TX 77098
of Publication: Houston, TX 77098 11. Known Bondholders,
PO Box 980985 Managing Editor: Mortagees, and Other Security
image processing functions is included that the capabilities are Houston, TX 77098 Jill Franklin Holders Owning or Holding 1
Contact Person: Mark Irgang PO Box 980985 Percent or More of Total
very deep and extensive, and because of Blender’s support for 713-589-3503 Houston, TX 77098 Amount of Bonds, Mortages,
HDR formats, such as OpenEXR, and its lack of limitation on reso- 8. Complete Mailing Address of or Other Securities: None
Headquarters of General 12. Tax Status: Has not Changed
lutions, it is squarely in the professional compositing camp, albeit Business Office of Publisher: During Preceding 12 Months
at the less-sophisticated end of the spectrum (as one would 13. Publication Title: Linux Journal 14. Issue Date: October
expect from a brand-new project). It is advancing quickly. In future 15. Extent and Nature of Circulation Average No. Copies Each Issue No. Copies of Single Issue
During Preceding 12 Months Published Nearest to Filing Date
versions, more user-friendly keying tools and color tools are a. Total Number of Copies:
planned, and hopefully there also will be more direct access to the (Net press run) 57,908 54,182
b. Paid and/or Requsted Circulation
translation and garbage matting functions, which at the moment (1) Paid/Requested Outside-County
are obtuse and inconvenient. Until such tools emerge, I highly Mail Subscriptions on Form 3541. 20,086 19,537
(2) Paid In-County Subscriptions
recommend that anyone wanting to use Blender as a workhorse Stated on Form 3541 0 0
compositor invest in a book that teaches how compositing works, (3) Sales Through Dealers and
Carriers, Street Vendors, Counter Sales,
both in theory and practice. The best available is The Art and and Other Non-USPS Paid Distribution 25,688 24,005
Science of Digital Compositing (The Morgan Kaufmann Series in c. Total Paid and/or Requested Circulation 45,774 43,542
d. Free Distribution Outside the Mail
Computer Graphics). (1) Outside-County as Stated
Using Blender for composite work has significant advantages as on Form 3541 810 780
(2) In-County as Stated
well, since it’s an integrated part of a 3-D content creation suite, on Form 3541 0 0
the particle systems, fluid systems, procedural textures and all the (3) Other Classes Mailed Through
the USPS 0 0
traditional 3-D modeling and animation tools are at the compositing e. Free Distribution Outside the Mail 2,343 1,720
system’s disposal, which is supremely useful for any number of f. Total Free Distribution 3,153 2,500
g. Total Distribution 48,927 46,042
highly complicated shots that normally would require using several h. Copies Not Distributed 8,981 8,140
i. Total 57,908 54,182
programs in conjunction to pull off correctly.
j. Percent Paid and/or Requested Circulation 94% 95%
Here’s hoping the Project Peach, the currently in-process sequel
PS Form 3526

94 | november 2007 w w w. l i n u x j o u r n a l . c o m
 /var/opinion

Is Hardware Catching Up to Java?

The demands of parallel processing may be
met more easily in a language we already know.

at the very least, keep them from advancing in BEA jRockit JVM, not the Sun JVM. See
speed as quickly as they might have if history www.intel.com/performance/server/
could have continued along its current path. xeon/java.htm for the chart and more
As a result, any single thread of execu- information. I suspect Intel used this JVM
tion would run faster on a speedy single-core because the BEA JVM employs garbage
CPU than it would on a slower core of a collection algorithms that are more efficient
multicore CPU. The seemingly obvious for use on multiple cores.
answer is to split up the task into multiple The fact that Intel used a non-Sun JVM
threads. Java is multithreaded by nature, makes this whole question of the future of Java
right? Therefore, all other things being on multicore CPUs interesting. I won’t discount
equal, a Java application should be a natural the possibility that Intel may have tuned its
for multicore CPU platforms, right? code to work best with the BEA JVM. But it is a
Not necessarily. If Java was a slam-dunk significant plus for Java that you can choose a
superior way to exploit parallel processing, I JVM best suited for the hardware you have.
would have posed this shift to Java as a predic- The big plus is that you still have to learn only
Nick Petreley, Editor in Chief tion, not a question. It’s not that simple. For one language, Java, and this does not limit
example, don’t let anyone tell you that Java was your choice of architecture or platforms. If you
I had a wonderful experience chatting with built from the ground up to exploit multiple run your application on a multicore platform,
the folks at RapidMind. The company pro- processors and cores. It ain’t so. Java’s famous you can choose between JVMs and JVM-specific
vides C/C++ programmers with a very clever garbage collection got in the way of parallel tuning parameters to squeeze extra perfor-
tool to help them exploit the parallel process- programming at first. Java versions as late as mance out of your application.
ing capability of multiple core CPUs. The 1.4 use a single-threaded garbage collector that Now, think a bit further into the future.
idea is to make parallel processing at least stalls your Java program when it runs, no matter Imagine parallelism showing up in a future
somewhat transparent to the programmer. how many CPUs you may have. generation of cell phones or other small
Programmers can use RapidMind to improve But Java’s multithreaded design more easily devices. What better language than a platform-
multicore performance of existing applica- exploits parallel processing than many other neutral one to take advantage of this future?
tions with minimal modifications to their languages, and it ended up lending itself to Some disclaimers are in order, though.
code and build new multicore applications improvements in garbage collection. JDK 5.0 First, I don’t think any tool or language will
without having to deal with all of the com- includes various tuning parameters that may soon make optimal programming for multi-
plexities of things like thread synchronization. minimize the impact of garbage collection on ple cores or CPUs totally transparent. Java
RapidMind and any other solutions like multiple cores or CPUs. It’s not perfect, and you has the potential to exploit multiple cores
it are invaluable and will remain so for a have to take into account the way your applica- with less impact to the programmer than
long, long time. C/C++ is an extremely well- tion is designed, but it is a vast improvement, many other languages, however, which is
entrenched legacy programming platform. and the improvement is made possible by the why I think Java’s future is getting even
I can’t help but wonder, however, if the fact that Java was well conceived from the start. more promising with this new hardware
trend toward increased parallel processing in Obviously, these features aren’t enough. trend. Second, although I think Java has
hardware will create a slow but steady shift IBM builds a lot of additional parallelism into the edge, other managed, multithreaded
toward inherently multithreaded, managed its WebSphere application server. In addition, languages will no doubt ride this same
programming platforms like Java. IBM and other researchers are working on a wave. Python, Ruby and your other favorite
The inherent trade-off to adding cores to Java-related language X10, which is designed language in this genre probably have an
a CPU poses an interesting problem for any specifically to exploit parallel processing (see equally bright future.I
programming platform. As you add cores, you x10.sourceforge.net/x10home.shtml).
increase power consumption, which generates It is also interesting to note that when Nicholas Petreley is Editor in Chief of Linux Journal and a former
more heat. One way to reduce heat is to lower Intel boasts about its quad-core performance programmer, teacher, analyst and consultant who has been
the processing speed of the individual cores, or on Java, its numbers are based on using the working with and writing about Linux for more than ten years.

96 | november 2007 w w w. l i n u x j o u r n a l . c o m
 Russ Barnard, President, FlapDaddy Productions

“Fanatical Support saved me ™

from my own mistake.”

“Not long ago, I reformatted one of our servers. Not until I was driving home did I learn that I brought our entire
site down in the process. I called my guy at Rackspace and he said, ‘We’re already on it.’ By the time I pulled
in the driveway, my site was back up. Now that’s Fanatical Support.”

Keeping little mistakes from causing big problems is one definition of Fanatical Support. What will yours be?

Watch Russ’s story at www.rackspace.com/fanatical

1-888-571-8976
 Affordable InfiniBand Solutions
4 Great Reasons to Call Microway NOW!
TriCom™ FasTree™
• DDR/SDR InfiniBand HCA • DDR InfiniBand switches
• "Switchless" serial console • Low latency, modular design
• NodeWatch web enabled • 24, 36 and 48 port building blocks
remote monitor and control
Headers to fan tach lines,
voltages, temperature probes
8051 BMC interface and
PS On/Off and MB reset
serial console switch
COM2
Internal connector

RJ45
RS-485/422
InfiniScope™
Daisy chain • Monitors ports on HCA’s and switches
connectors • Provides real time BW diagnostics
Mellanox™ InfiniHost III
InfiniBand HCA
• Finds switch and cable faults
InfiniBand
connector • Lane 15 interface
• Logs all IB errors

ServaStor™
• Extensible IB based storage
building blocks
• Redundant and scalable
• Parallel file systems
• Open source software
• On-line capacity expansion
• RAID 0,1,1E, 3, 5, 6, 10, 50

Upgrade your current cluster, or let us design your

next one using Microway InfiniBand Solutions.

To speak to an HPC expert

call 508 746-7341 and ask
for technical sales or email
[email protected]
www.microway.com