Building Clusters With FreeBSD

Brooks Davis
The Aerospace Corporation
<brooks@{aero,freebsd}.org>
March 8, 2007
http://people.freebsd.org/~brooks/pubs/asiabsdcon2007/

© 2006-2007 The Aerospace Corporation

 Tutorial Outline
● Introductions
● Overview of Fellowship
● Cluster Architecture Issues
● Operational Issues
● Thoughts on a Second Cluster
● FreeBSD specifics
 Introductions
● Name
● Affiliation
● Interest in clusters
Overview of Fellowship
 Overview of Fellowship
● The Aerospace Corporation's corporate,
unclassified computing cluster
● Designed to be a general purpose cluster
– Run a wide variety of applications
– Growth over time
– Remote access for maintainability
● Gaining experience with clusters was a goal
● In production since 2001
● >100 users
 Overview of Fellowship
Software
● FreeBSD 6.1
● Sun Grid Engine (SGE) scheduler
● Ganglia cluster monitor
● Nagios network monitor
 Overview of Fellowship
Hardware
● 353 dual-processor nodes
– 64 Intel Xeon nodes
– 289 Opteron nodes (152 dual-core)
● 3 TB shared (NFS) disk
● >60TB total storage
● 700GB RAM
● Gigabit Ethernet
– 32 nodes also have 2Gbps Myrinet
 Overview of Fellowship
Facilities
● ~80KVA power draw
– Average US house can draw 40KVA max
● 273 kBTU/hr ~= 22Tons of refrigeration
● 600 sq ft floor space
– Excluding HVAC and power distribution
 Overview of Fellowship
Network Topology
r01n01
r01n02
fellowship r01n03
...
frodo
r02n01
gamgee 10.5.0.0/16 r02n02
Aerospace
Network r02n03
arwen Cat6509
...
elrond r03n01
r03n01
moria r03n02
r03n02
r03n03
r03n03
...
...
 Cluster Architecture Issues
● Architecture matters
– Mistakes are compounded when you buy
hundreds of machines
● Have a requirements process
– What are your goals?
– What can you afford?
● Upfront
● Ongoing
 Cluster Architecture Issues
● Operating System
● Processor Architecture
● Network Interconnect
● Storage
● Form Factor
● Facilities
● Scheduler
 Cluster Architecture Issues
Slide Format
● Trade offs and Considerations
– The trade space and other things to considers
● Options
– Concrete options
● What we did on Fellowship
● How it worked out
 Operating System
Trade offs and Considerations
● Cost: Licensing, Support

● Performance: Overhead, Driver quality

● Hardware Support: Processor, Network,

Storage
● Administration: Upgrade/patch process,

software installation and management

● Staff experience: software porting,

debugging, modification, scripting

 Operating System
Options
● Linux

– General purpose distros: Debian, Fedora, Red

Hat, SuSE, Ubuntu, etc.
– Cluster kits: Rocks, OSCAR
– Vendor specific: Scyld
● BSD: FreeBSD, NetBSD, OpenBSD
● MacOS/Darwin
● Commercial Unix: Solaris, AIX, HPUX,
Tru64
● Windows
 Operating System
What we did on Fellowship
● FreeBSD

– Started with 4.x

– Moved to 6.x
How it worked
● Netboot works well

● Linux emulation supports commercial code

(Mathematica, Matlab)
● No system scope threads in 4.x (fixed in 5.x)

● Had to port SGE, Ganglia, OpenMPI

● No parallel debugger
 Processor Architecture
Trade offs and Considerations
● Cost

● Power consumption

● Heat production

● Performance: Integer, floating point, cache

size and latency, memory bandwidth and

latency, addressable memory
● Software Support: Operating system,

hardware drivers, applications (libraries),

development tools
 Processor Architecture
Options
● IA32 (i386): AMD, Intel, Transmeta, Via

● AMD64 (EM64T): AMD, Intel

● IA64 (Itanium)

● SPARC

● PowerPC

● Power

● Alpha

● MIPS

● ARM
 Processor Architecture
What we did on fellowship
● Intel Pentium III's for the first 86

● Intel Xeons for the next 76

● AMD Opterons for the most recent

purchases (169)
● Retired Pentium III's this year

How it worked
● Pentium III's gave good service

● Xeons and Opterons performing well

● Considering 64-bit mode for the future

● Looking at Intel Woodcrest CPUs

 Network Interconnects
Trade offs and Considerations
● Cost: NIC, cable, switch ports

● Performance: throughput, latency

● Form factor: cable management and

termination
● Standardization: commodity vs proprietary

● Available switches: size, inter-switch links

● Separation of different types of traffic

 Network Interconnects
Options
● 10/100 Ethernet

● Gigabit Ethernet

● 10 Gigabit Ethernet: fast

● Infiniband: fast, low latency

● 10 Gb Myrinet: fast, low latency

● Others: Dolphin, Fiber Channel

 Network Interconnects
What we did on Fellowship
● Gigabit Ethernet

● One rack of 2Gbps Myrinet nodes

How it worked
● Gigabit Ethernet is now the default option for

clusters
● Fast enough for most of our applications

● Some applications would like lower latency

● Looking at 10GbE and 10Gb Myrinet

 Storage
Trade offs and Considerations
● Cost

● Capacity

● Throughput

● Latency

● Locality

● Scalability

● Manageability

● Redundancy
 Storage
Options
● Local Disk

● Protocol Based Network Storage: host or

NAS appliance based

● Storage Area Network

● Clustered Storage
 Storage
What we did on fellowship
● Host based NFS for home directories, node

roots, and some software

● Local disks for scratch and swap

● Moved home directories to a Netapp in 2005

How it worked
● NFS is scaling fine so far

● Enhanced Warner Losh's diskprep script to

keep disk layouts up to date

● Users keep filling the local disks

● Disk failures are a problem

 Form Factor
Trade offs and Considerations
● Cost

● Maximum performance

● Maintainability

● Cooling

● Peripheral options

● Volume (floor space)

● Looks
 Form Factor
Options
● PCs on shelves

● Rackmount system

– Cabinets
– 4-post racks
– 2-post racks
● Blades
 Form Factor
What we did on fellowship
● 1U nodes in 2-post racks

● Core equipment in short 4-post racks

● 6 inch wide vertical cable management with

direct runs from the switch in first row

● Moved to 10 inch wide vertical management

in second row and patch panels in both rows

● Now installing new core equipment in

cabinets
Form Factor
 Form Factor
How it worked
● Node racks are accessible and fairly clean

looking
● Patch panels, 10 inch cable management,

and some custom cable lengths helped

● Short 4-post racks didn't work well for real

servers
● Watch out for heavy equipment!
 Facilities
Trade offs and Considerations
● Cost: space, equipment, installation

● Construction time

● Reliability
 Facilities
Options
● Plug it in and hope

● Convert a space (office, store room, etc)

● Build or acquire a real machine room

● Use an old mainframe room

Facilities
 Facilities
What we did on Fellowship
● Built the cluster in our existing 15,000 sq ft.

underground machine room

– 500KVA building UPS and two layers of backup
generators
● New UPS and power distribution units
(PDUs) being installed for expansion
 Facilities
How it worked
● Good space with plenty of cooling

● Power was initially adequate, but is

becoming limited
– Adding a new UPS and PDUs
● Cooling issues with new UPS
● Remote access means we don't have to
spend much time there
 Scheduling Scheme
Trade offs and Considerations
● Cost

● Efficiency

● Support of policies

● Fit to job mix

● User expectations
 Scheduling Scheme
Options
● No scheduler

● Custom or application specific scheduler

● Batch job system

● Time sharing
 Scheduling Scheme
What we did on Fellowship
● None initially

● Tried OpenPBS (not stable 4 years ago, no

experience since)
● Ported Sun Grid Engine (SGE) 5.3 with help

from Ron Chen

● Switched to SGE 6 and mandated use in

January
 Scheduling Scheme
How it worked
● Voluntary adoption was poor

● Forced adoption has gone well

● Users have preconceived notions of

computers that don't fit reality with batch

schedulers
● We have modified SGE to add features

missing from the FreeBSD port with good

success
 Operational Issues
● Building, Refresh and Upgrade Cycle
● User Configuration Management
● System Configuration management
● Monitoring
● Inventory Management
● Disaster Recovery
Initial Build, Refresh and Major
Upgrade Cycle
Trade offs and Considerations
● Startup cost

● Ongoing cost

● Homogeneity vs Heterogeneity

● Gradual migration vs abrupt transitions

Initial Build, Refresh and Major
Upgrade Cycle
Options
● Build

– Build all at once

– Gradual buildup
● Refresh
– Build a new cluster before retirement
– Build a new cluster in the same location
– Replace parts over time
● Upgrades
– Upgrade everything at once
– Partion and gradually upgrade
– Never upgrade
Initial Build, Refresh and Major
Upgrade Cycle
What we did on Fellowship
● Build

– Gradual buildup of nodes

– Periodic purchase of new core systems for
expansion and replacement
● Refresh
– Replaced PIII's this year
– Xeons to be replaced next year if we don't
expand to a third row
● Upgrades
– Minor OS upgrades in place
– FreeBSD and SGE 6 by partitioning
Initial Build, Refresh and Major
Upgrade
How it worked
Cycle
● Build
– Most of our apps don't care
– Different machines had different exposed serial
ports which caused a problem for serial
consoles
● Refresh
– Rapid failures of Pentium III's were unexpected
● Major Upgrades
– Partitioning allowed a gradual transition
– New machines offered incentive to move
– Node locked SSH keys and licenses caused
problems
 System Testing
Trade offs and Considerations
● Need to validate system stability and

performance
– LLNL says: “bad performance is a bug”
● “Bad batches” of hardware happen
● Lots of hardware means the unlikely is much
more common
 System Testing
Options
● Leave it to the vendor

● Have a burn-in period

– No user access
– Limited user access
● Periodic testing
 System Testing
What we did on Fellowship
● Vendor burn in

– Increasingly strict requirements to ship

● Let users decide where to run (prior to
mandatory scheduling)
● Scheduler group of nodes needing testing
● Working on building up a set of performance
and stress tests
 System Testing
How it worked
● Ad hoc testing makes problems surprising

too often
● Users find too many hardware issues before

we do
● Group of nodes is easy to administer
 System Configuration
Management
Trade offs and Considerations
● Network Scalability

● Administrator Scalability

● Packages vs custom builds

● Upgrading system images vs new, clean

images
 System Configuration
Management
Options
● Maintaining individual nodes

● Push images to nodes

● Network booted with shared images

– Read only
– Copy-on-write
 System Configuration
Management
What we did on Fellowship
● PXE boot node images with automatic
formatting of local disks for swap and
scratch
● Upgraded copies of the image in 4.x

● Building new images for each upgrade in 6.x

How it worked
● Great overall

● A package build system to help keep

frontend and nodes in sync would be nice

● Network bottle neck does not appear to be a

problem at this point

 User Configuration
Management
Trade offs and Considerations
● Maintainability

● User freedom and comfort

● Number of supported shells

 User Configuration
Management
Options
● Make users handle it

● Use /etc/skel to provide defaults and have

users do updates
● Use a centrally located file that users source

● Don't let users do anything

 User Configuration
Management
What we did on Fellowship
● /etc/skel defaults plus users updates to start

● Added a central script recently

– This script uses an sh script and some wrapper

scripts to work with both sh and csh style shells
● Planning a manual update
How it worked
● Bumpy, but improving with the central script
 Monitoring
Trade offs and Considerations
● Cost

● Functionality

● Flexibility

● Status vs alarms
 Monitoring
Options
● Cluster management systems

● Commercial network management systems:

Tivoli, OpenView
● Open Source system monitoring packages:

Big Sister, Ganglia, Nagios

● Most schedulers

● SNMP
 Monitoring
What we did on Fellowship
● Ganglia early on

● Added Nagios recently

● SGE

How it worked
● Ganglia provides very user friendly output

– Rewrote most of FreeBSD support

● Nagios working well
● Finding SGE increasingly useful
 Disaster Recovery
Trade offs and Considerations
● Cost up front

● Cost of recovery

● Time to recovery

● From what type of disaster

– Hardware failure
– Loss of building
– Data contamination/infection/hacking
 Disaster Recovery
Options
● Do nothing

● Local backups

● Off site backups

● Geographically redundant clusters

– Transparent access to multiple clusters

 Disaster Recovery
What we did on Fellowship
● Local backups (Bacula, formerly AMANDA)

● Working toward off site backups

How it worked
● No disasters yet

● Local backups are inadequate

● Looking at a second cluster

● Investigating transparent resource discovery

and access
 Other Issues
● Virtualization
● System Naming and Addressing
● User Access
● Administrator Access
● User Training and Support
● Inventory Management
Thoughts on a Second Cluster
● We are planning to build a second, similar
cluster on the east coast
● Looking at blades for density and
maintenance
● Interested in higher speed, lower latency
interconnects for applications which can use
them
● Considering a completely diskless approach
with clustered storage to improve
maintainability and scalability
 FreeBSD Specifics
● Diskless booting
– Image creation
– Disk initialization
● Using ports on a Cluster
● Ganglia demo
● SGE installation and configuration demo
 Diskless Booting:
Image Creation
● Hacked copy of nanobsd Makefile
– Removed flash image support
– Added ability to create extra directories for use
as mount points
– Build a list of ports in the directory via chroot
● Ports directory created with portsnap
● Ports are built using portinstall in a chroot
● Mount linprocfs before every chroot and unmount it
afterward
● Distfile pre-staging is supported for non-redistributable
distfiles and faster rebuilds
● Packages are also supported
● DESTDIR support in ports will eventually
make this obsolete
 Diskless Booting:
Image Creation
TODO
● Switch to nanobsd scripts (in place of

obsolete Makefiles)
● Handle sudoers file in images

– Copy on in place after install, extend

rc.initdiskless /conf support to /usr/local/etc, or
add the ability to override in port
● Find a way to keep packages in sync
between nodes and front end systems
 Diskless Booting
Startup Process
● PXE boot with NFS root
● /etc/rc.initdiskless initializes /etc from data in
/conf (mounted from /../conf to allow sharing)
– /conf/base/etc remounts /etc
– /conf/default/etc includes rc.conf which simply
sources rc.conf.{default,bcast,ipaddr} allowing
configuration to live in the right place
● /etc/rc.d/diskprep creates swap, /tmp, and
/var and labels them to fstab stays
consistant reguardless of disk configuration
● Normal boot from this point on
 Diskless Booting:
Disk Initalization
● Use sysutils/diskprep port (modified version
of Warner Losh's tool for embedded
deployments
– If the right GEOM volume label doesn't exist,
reconfigure the disk
● Could be improved
– Reboot during initalization is often fatel
– Better control of fsck at boot would be useful
● Option to newfs file systems who's contents we don't
care about
– Alternate superblock printout in newfs too noisy
 Using Ports on a Cluster
● Very good for languages and cluster tools
● Unusable for MPI ports due to the need for
different ones with different compilers
– Need a bsd.mpi.mk
● Mixed for libraries
– Some are fine with one compiler but others
could benefit from more than on version,
particularly Fortran code
● Hard to keep nodes and front ends in sync
– Need an SGE package cluster :)
 Using Ports on a Cluster
Useful Ports
● lang/gcc*, lang/icc, lang/ifc, etc.

● net-mgmt/nagios

● sysutils/diskprep

● sysutils/ganglia-monitor-core

● sysutils/ganglia-webfrontend

● sysutils/sge
Ganglia Demo
SGE Demo

Configuration Demo
Background Slides
 Aside: Virtualization
● Virtualization presents another paradigm for
clusters
● Multiple operating systems can be supported
allowing applications to run where the work
best
● Migration of jobs can allow for simplified
maintenance
● Time sharing of machines is more practical
than with normal batch systems
 System Naming and
Addressing
Trade offs and Considerations
● Ease of accessing resources

● Easy of physically locating equipment from a

name or address
● Address allocation efficiency
 System Naming and
Addressing
Options
● Naming

– By address (e.g. 10.5.2.10-node)

– By location (e.g. Rack2node10)
– Sets of things (elements, Opera Singers,
FreeBSD committers, etc)
● Addressing
– By location vs arbitrary
– Public vs private
– Packed vs sparse
– IPv6
 System Naming and
Addressing
What we did on Fellowship
● Naming

– Core systems are Lord of the Rings characters

– Nodes by rack and unit (r02n10)
– Cluster DNS zone (fellow.aero.org)
● Addressing
– Private network 10.5/16
– Node addresses 10.5.rack.unit
– NAT for downloads
 System Naming and
Addressing
How it worked
● Has forced some application design issues

(no servers on nodes)

● Good demonstration of the issues of a

private address space cluster

● IPv6 would be nice
 User Access
Trade offs and Considerations
● Ease of use

● User familiarity/comfort

● Control of resources
 User Access
Options
● Shell on a frontend machine

● Direct access to nodes

● Single system image

● Desktop integration

● Web based portals

● Application integration
 User Access
What we did on Fellowship
● SSH to frontend where users edit and compile

code and submit jobs to the scheduler

● Working on grid based solution for easier access

– Web portals
– Integration with clients
How it worked
● Many users don't really get the command prompt

thanks to Microsoft and Apple

● Some initial resistance to all access via SSH
 Administrator Access
Trade offs and Considerations
● Cost

● Effectiveness of out of band access

● Frequency of use
 Administrator Access
Options
● SSH to machines

● IPMI

● Serial consoles

● Local KVM switches

● Remote KVM switches

 Administrator Access
What we did on Fellowship
● SSH as primary access

● Serial console to nodes and core systems

initially
● Local KVM access to core systems

● Upgraded to remote KVM access

 Administrator Access
How it worked
● SSH is great when it works, but high latency

with many short connections

● We have abandoned serial consoles as too

expensive for the infrequent use

● Remove KVM access is very nice

● All devices except the Cisco 6509 have

remote power control

● Would like to investigate IPMI
 Inventory Management
Trade offs and Considerations
● Need to know what hardware and software

is where
● Need history of success/failure to detect

buggy hardware
 Inventory Management
Options
● Ad hoc tracking

● Wiki

● Property database

● Some request tracking systems

 Inventory Management
What we did on Fellowship
● Ad hoc then some wiki pages

● Investigating a database solution

How it worked
● Sort of works, but things get lost or forgotten
 Disclaimer
● All trademarks, service marks, and trade
names are the property of their respective
owners.