2009 cluster user training

Cluster User Training
From Bash to parallel jobs under SGE in one terrifying
hour
Christopher Dwan, Bioteam
First delivered at IICB, Kolkata, India
December 14, 2009

Unix command line essentials
“pwd”: Print current working directory (where am I?)
“ls”: File listing
“ls –l”: Detailed file listing, including permissions
“cd”: Change directory
“chmod”: Change file permissions
“echo”: Print something to the terminal
“more”: Show contents of a text file
“pico” Text editor
“man”: Documentation on any Unix command:

Basic Unix Exercises: Making directories
What directory am I currently in?
Remora:~ cdwan$ pwd
/Users/cdwan/
Create a directory named “test_1”
Remora:~ cdwan$ mkdir test_1
Remora:~ cdwan$ ls
test_1
Change into that directory, and verify that we are there.
Remora:~ cdwan$ cd test_1
remora:test_1 cdwan$ pwd
/Users/cdwan/test_1

More basic Unix
Return to your home directory:
“cd” with no arguments
Exit the session:
“exit”

File editing.
“The best script editor” is the subject of an ongoing religious war in
technical circles
You should use the tool that does not get in your way.
“vi”: lightweight, complex, powerful, difficult to use
“emacs”: heavyweight, complex, powerful, difficult to use
“pico”: Possibly the simplest editor to use
To edit a file: “pico filename”

Hello world in ‘bash’
“Bash” is a shell scripting language.
– It is the default scripting language that you have at the terminal.
– I.e: You are already using it.
– We will take this command:
remora:test_1 cdwan$ echo "hello world"
hello world
And create a wrapper script to do the same thing:
remora:test_1 cdwan$ pico hello.sh
remora:test_1 cdwan$ chmod +x hello.sh
remora:test_1 cdwan$ ./hello.sh
hello world

Running a set of bash commands
Using pico, create a file named “hello.sh” containing a single line:
echo "hello world"
Exit pico. Verify the contents of the file:
remora:ex_1 cdwan$ more hello.sh
echo "hello world"
Then invoke it using the ‘bash’ interpreter:
remora:ex_1 cdwan$ bash hello.sh
hello world

Hello world script
#!/bin/bash
echo "hello world”
The “#!” line tells the system to automatically run it using bash
Ctrl-O: Save the file
Ctrl-X: Exit pico

File permissions
Files have properties:
– Read, Write, Execute
– Three different types of user: “owner”, “group”, “everyone”
To take a script you have written and turn it into an executable program,
run “chmod +x” on it.
remora:test_1 cdwan$ ls -l hello.sh
-rw-r--r-- 1 cdwan staff 32 Dec 14 21:56 hello.sh
remora:test_1 cdwan$ ls -l hello.sh
-rwxr-xr-x 1 cdwan staff 32 Dec 14 21:56 hello.sh

Execute the hello world script
remora:test_1 cdwan$ ./hello.sh
hello world

Most useful SGE commands
• qsub / qdel
– Submit jobs & delete jobs
• qstat & qhost
– Status info for queues, hosts and jobs
• qacct
– Summary info and reports on completed job
• qrsh
– Get an interactive shell on a cluster node
– Quickly run a command on a remote host
• qmon
– Launch the X11 GUI interface
Please do not copy, put online
or redistribute
info@bioteam.net

Interactive Sessions
To generate an interactive session, scheduled on any node:
“qlogin”
applecluster:~ cluster$ qlogin
Your job 145 ("QLOGIN") has been submitted
waiting for interactive job to be scheduled ...
Your interactive job 145 has been successfully scheduled.
Establishing /common/node/ssh_wrapper session to host node002.cluster.private ...
The authenticity of host '[node002.cluster.private]:50726 ([192.168.2.2]:50726)'
can't be established.
RSA key fingerprint is a7:02:43:23:b6:ee:07:a8:0f:2b:6c:25:8a:3c:93:2b.
Are you sure you want to continue connecting (yes/no)? yes
Warning: Permanently added '[node002.cluster.private]:50726,[192.168.2.2]:50726'
(RSA) to the list of known hosts.
Last login: Thu Dec 3 09:55:42 2009 from portal2net.cluster.private
node002:~ cluster$
all.q@node002.cluster.private BIP 1/8 0.00 darwin-x86
145 0.55500 QLOGIN cluster r 12/15/2009 09:15:08

Requesting the whole node
qlogin -pe threaded 8
all.q@node014.cluster.private BIP 8/8 0.00 darwin-x86
146 0.55500 QLOGIN cluster r 12/15/2009 09:34:58 8
We request a parallel environent called “threaded” (note, this PE does not
exist by default in SGE – we create it in iNquiry)
We request 8 slots within that environment
Now, no other jobs will be scheduled to your node while that login is in
place.

Most basic job example
qsub –b y /bin/hostname
You will see two new files in your home directory:
hostname.oYYY
hostname.eYYY
YYY is the job id provided by the queuing system.
These are the standard output and standard error files from running
/bin/hostname on one of the nodes.
Argument “-b y” indicates that this is a a compiled binary. SGE will not
try to parse the input, but merely run it.

Creating the sleeper script
#!/bin/bash
echo “hello world”
sleep 60
hostname
Then run like this:
remora:test_1 cdwan$ cp hello.sh sleeper.sh
remora:test_1 cdwan$ pico sleeper.sh
remora:test_1 cdwan$ ./sleeper.sh
hello world
remora.local

Submitting the sleeper script
genesis2:example bioteam$ qsub -cwd –S /bin/bash sleeper.sh
Your job 217 ("sleeper.sh") has been submitted
genesis2:example bioteam$ qstat
job-ID prior name user state submit/start at
queue slots ja-task-ID
-----------------------------------------------------------------
-------
217 0.55500 sleeper.sh bioteam r 12/14/2009 12:00:46
all.q@node004.cluster.private 1

Adding arguments directly into the script
#!/bin/bash
#$ -S /bin/bash
#$ -cwd
echo "hello world”
sleep 60
hostname
Any comment that starts with “#$” is interpreted as an argument to
qsub.
In case of conflict, the command line wins.

More SGE commands
• “qstat –f”: Show all queues, even the empty ones
• “qstat –u *”: Show jobs from all users
• “qstat –f –u *”: Both all queues and uses
• “qdel job_id”: Delete a particular job
• “qdel –u cdwan”: Delete all jobs run by user cdwan

Giving your job a name
#!/bin/bash
#$ -S /bin/bash
#$ -cwd
#$ -N sleeper
echo ”Hello world”
sleep 60
hostname
genesis2:demo bioteam$ qsub sleeper.sh
Your job 222 ("sleeper") has been submitted

Resource Requirements
genesis2:demo bioteam$ qsub -l arch=solaris sleeper.sh
Your job 225 ("sleeper") has been submitted
genesis2:demo bioteam$ qstat
job-ID prior name user state
----------------------------------------------
225 0.00000 sleeper bioteam qw
We specify a resource requirement that cannot be met (there are no solaris
machines in the cluster)
Qstat –j 225 tells the story:
(-l arch=solaris) cannot run at host "node008.cluster.private" because it offers only
hl:arch=darwin-ppc

Environment variables from SGE
SGE sets several variables in the script for you.
– JOB_ID numerical ID of the job
#!/bin/bash
#$ -S /bin/bash
#$ -cwd
#$ -N sleeper
echo "My job id is " $JOB_ID
sleep 60
genesis2:demo bioteam$ more sleeper.o221
My job id is 221

Job dependencies
genesis2:demo bioteam$ qsub -N ”primary” sleeper.sh
genesis2:demo bioteam$ qsub -hold_jid primary -N
"secondary" sleeper.sh
Your job 224 ("secondary") has been submitted
genesis2:demo bioteam$ qstat
job-ID prior name user state
-----------------------------------------------
223 0.55500 primary bioteam r
224 0.00000 secondary bioteam hqw

Redirecting output
#!/bin/bash
#$ -S /bin/bash
#$ -cwd
#$ -o task.out
#$ -e task.err
echo "Job ID is " $JOB_ID

Task arrays
#!/bin/bash
#$ -S /bin/bash
#$ -cwd
#$ -N task_array
#$ -o task.out
#$ -e task.err
echo "Job ID is " $JOB_ID “ Task is “ $SGE_TASK_ID

Task arrays
genesis2:example_2 bioteam$ qsub -t 1-10 task.sh
Your job-array 228.1-10:1 ("task_array") has been
submitted
genesis2:example_2 bioteam$ more task.out
Job ID is 228 Task ID is 10
• …

Perl instead of bash
My “hello world” script, using Perl instead of Bash:
remora:~ cdwan$ more hello_world.pl
#!/usr/bin/perl
#! –S /usr/bin/perl
sleep(60);
print "Hello worldn";
Submitted to the queuing system:
qsub hello_world.pl

PERFORMANCE TUNING /
PARALLELIZATION

Parallel Jobs
• A parallel job runs simultaneously across multiple servers
– Biggest SGE job I’ve heard of: Single application running across 63,000
CPU cores: TACC “Ranger” Cluster in Texas
– Distinction with ‘batches’ of processes that include many tasks to be
done in any order
APPLICATION

cdwan@bioteam.net
Parallel computing 101
APPLICATION
High Performance
Computing
High Throughput
Computing (farm)

Batch Jobs
Copyright 2006, The BioTeam
32 Not for
Redistribution
http://bioteam.net
Private Ethernet Network
“Public” Ethernet Network
•Independent applications running at the same time
•Many jobs (batch)
•Maximum efficiency, simple to write

Tightly Coupled / Parallel
Copyright 2006, The BioTeam
33 Not for
Redistribution
http://bioteam.net
One parallel application running over the entire cluster
Private Ethernet Network
“Public” Ethernet Network
Application
•One job, where response time is important.
•Overall efficiency is lower
•Scalability is hard

Amdahl’s Law
Maximum expected speedup for
parallelizing any task
– Serial portion (non parallelizable)
– Parallel portion (can be parallelized)
Additionally:
– Cost associated with using more
machines (startup, teardown)
– At least, scheduling. Possibly some
other factors like communication
– Communication scales with number
of processes
Important to note:
Re-stating the problem can radically
alter the serial / parallel ratio

Network Latency
• Latency:
– Time to initiate communication
• Throughput:
– Data rate once communication is established
• Gigabit Ethernet:
– Latencies: ~100ms
– Throughputs up to 80% of wire speed (800Mb/sec)
– $10 / network port
• Myranet / Infiniband:
– Latencies: ~3ms
– Throughput: 80% of wire speed
– $800 / network port

Parallel Jobs
Many different software implementations are used to support
parallel tasks:
– MPICH
– LAM-MPI
– OpenMPI
– PVM
– LINDA
No magic involved
– Requires work
– Your application must support parallel methods

Submitting a standalone MPI job
Build the code with a particular version of MPI:
genesis2:examples bioteam$ pwd
/common/mpich/ch_p4/examples
genesis2:examples root# which mpicc
/common/mpich/ch_p4/bin/mpicc
genesis2:examples root# mpicc cpi.c
genesis2:examples root# mpicc -o cpi cpi.o
Run without any MPI framework:
genesis2:examples root# ./cpi
Process 0 on genesis2.mit.edu
pi is approximately 3.1416009869231254, Error is
0.0000083333333323
wall clock time = 0.000214

Submitting a standalone MPI job (no SGE)
MPI needs to know which hosts to use: We create a hosts file which simply lists
the machine
genesis2:~ bioteam$ more hosts_file
node001
node002
Node003
node004
Then start the job using ‘mpirun’
genesis2:~ bioteam$ mpirun -machinefile hosts_file
-np 4 /common/mpich/ch_p4/examples/cpi
Process 0 on genesis2.mit.edu
Process 2 on node002.cluster.private
pi is approximately 3.1416009869231249, Error is 0.0000083333333318

Critical Notes
In order to have MPICH jobs work reliably, you need to compile
and run them with the same version of MPICH.
/common/mpich
/common/mpich2
/common/mpich2-64
All user account issues must be in order for this to work.
– Password free ssh in particular
If the application does not work from the command line, SGE will
not help.

Loose Integration with SGE
Loose Integration
– Grid Engine used for:
• Picking when the job runs
• Picking where the job runs
• Generating the custom machine file
– Grid Engine does not:
• Launch or control the parallel job itself
• Track resource consumption or child processes
Advantages of loose integration
– Easy to set up
– Can trivially support almost any parallel application technology
Disadvantages of loose integration
– Grid Engine can’t track resource consumption
– Grid Engine must “trust” the parallel app to honor the custom hostfile
– Grid Engine can not kill runaway jobs

Tight integration with SGE
Tight Integration
– Grid Engine handles all aspects of parallel job operation from start to
finish
– Includes spawning and controlling all parallel processes
Tight integration advantages:
– Grid Engine remains in control
– Resource usage accurately tracked
– Standard commands like “qdel” will work
• Child tasks will not be forgotten about or left untouched
Tight Integration disadvantages:
– Can be really hard to implement
– Makes job debugging and troubleshooting harder
– May be application specific

Running an mpich job with loose SGE
integration
Step one: Job must work without SGE.
– Until you can demonstrate a running job using a host file and manual
start up, there is no point to involving SGE
Step two: Create a wrapper script to allow SGE to define the list of
hosts and the number of tasks
Step three: Submit that wrapper script into a ‘parallel
environment’
– Parallel environment manages all the host list details for you.

MPICH Wrapper for CPI
A trivial MPICH wrapper for Grid Engine:
#!/bin/bash
## ---- EMBEDDED SGE ARGUMENTS ----
#$ -N MPI_Job
#$ -pe mpich 4
#$ -cwd
#$ -S /bin/bash
## ------------------------------------
MPIRUN=/common/mpich/ch_p4/bin/mpirun
PROGRAM=/common/mpich/ch_p4/examples/cpi
export RSHCOMMAND=/usr/bin/ssh
echo "I got $NSLOTS slots to run on!"
$MPIRUN -np $NSLOTS -machinefile $TMPDIR/machines $PROGRAM

Job Execution
Submit just like any other SGE job:
[genesis2:~] bioteam% qsub submit_cpi
Your job 234 ("MPI_Job") has been submitted
Output files generated:
[genesis2:~] bioteam% ls -l *234
-rw-r--r-- 1 bioteam admin 185 Dec 15 20:55 MPI_Job.e234
-rw-r--r-- 1 bioteam admin 120 Dec 15 20:55 MPI_Job.o234
-rw-r--r-- 1 bioteam admin 52 Dec 15 20:55 MPI_Job.pe234
-rw-r--r-- 1 bioteam admin 104 Dec 15 20:55 MPI_Job.po234

Output
[genesis2:~] bioteam% more MPI_Job.o234
I got 5 slots to run on!
pi is approximately 3.1416009869231245, Error is
0.0000083333333314
[genesis2:~] bioteam% more MPI_Job.e234

Parallel Environment Usage
• “qsub -pe mpich 4 ./my-mpich-job.sh”
• “qsub -pe mpich 4-10 ./my-mpich-job.sh”
• “qsub -pe lam-loose 3 ./my-lam-job.sh”

Behind the Scenes: Parallel Environment
Config
genesis2:~ bioteam$ qconf -sp mpich
pe_name mpich
slots 512
user_lists NONE
xuser_lists NONE
start_proc_args /common/sge/mpi/startmpi.sh $pe_hostfile
stop_proc_args /common/sge/mpi/stopmpi.sh
allocation_rule $fill_up
control_slaves FALSE
job_is_first_task TRUE
urgency_slots min

Behind the scenes: MPICH
The “startmpi.sh” script is run before job launches and
creates custom machine file
The user job script gets date required by ‘mpirun’ from
environment variables:
$NODES, $TEMPDIR/machines, etc.
The “stopmpi.sh” script is just a placeholder
Does not really do anything (no need yet)

Behind the scenes: LAM-MPI
• Just like MPICH
• But 2 additions:
– The “lamstart.sh” script launches LAMBOOT
– The “lamstop.sh” script executes LAMHALT at job termination
• In an example configuration, lamboot is started this way:
– lamboot -v -ssi boot rsh -ssi rsh_agent "ssh -x
-q" $TMPDIR/machines

Behind the scenes: LAM-MPI
• A trivial LAM-MPI wrapper for Grid Engine:
#!/bin/sh
MPIRUN=“/common/lam/bin/mpirun”
#$ -N MPI_Job
#$ -pe lammpi 3-5
#$ -cwd
## ------------------------------------
echo "I have $NSLOTS slots to run on!"
$MPIRUN C ./mpi-program

OpenMPI
In absence of specific requirements, a great choice
Works well over Gigabit Ethernet
Trivial to achieve tight SGE integration
Recent personal experience:
– Out of the box: ‘cpi.c’ on 1024 CPUs
– Out of the box: heavyweight genome analysis pipeline on 650 Nehalem
cores

Behind the scenes: OpenMPI
OpenMPI 1.2.x natively supports automatic tight SGE integration
– Build from source with “--enable-sge”
– mpirun -np $NSLOTS /path-to-my-parallel-app
OpenMPI PE config:
pe_name openmpi
slots 4
user_lists NONE
xuser_lists NONE
start_proc_args /bin/true
stop_proc_args /bin/true
allocation_rule $round_robin
control_slaves TRUE
job_is_first_task FALSE
urgency_slots min

OpenMPI Job Script
#!/bin/sh
#$ -N MPI_Job
#$ -pe openmpi 3-5
#$ -cwd
## ------------------------------------
echo "I got $NSLOTS slots to run on!"
mpirun -np $NSLOTS ./my-mpi-program

Application profiling
• Most basic: System Monitoring
– ‘top’
– Ganglia
• Apple ‘shark’ tools
• Deep understanding of code.

Tuning parallel jobs
• Round Robin:
– Jobs are distributed to as many nodes as possible
– Good for tasks where memory may be the bottleneck
• Fill up:
– Jobs are packed onto as few nodes as possible
– Good for jobs where interprocess communications may be the
bottleneck
• Single chassis
– “threaded” environment from earlier sessions
– For multi-threaded programs (BLAST)

2009 cluster user training

Recommended

More Related Content

What's hot (20)

Similar to 2009 cluster user training (20)

More from Chris Dwan (20)

Recently uploaded (20)

2009 cluster user training