node_local_scheduling

This is an old revision of the document!


Node-local scheduling

There are some use cases, where you would want to simply request a full cluster node from the LSF batch system and then run many (e.g. much more than 64) smaller (e.g. only a fragment of the total job runtime) tasks on this full node. Then of course you will need some local scheduling on this node to ensure proper utilization of all cores.

To accomplish this, we suggest you use the GNU Parallel program. The program is installed to /cluster/bin, but you can also simply load the modulefile software/gnu_parallel so that you can also access its man page.

For more documentation on how to use GNU Parallel, please read man parallel and man parallel_tutorial, where you'll find a great number of examples and explanations.

Let's say we have a number of input data files that contain differing parameters that are going to be processed independently by our program:

$ ls data_*.in
data_001.in
data_002.in
[...]
data_149.in
data_150.in
$ cat data_001.in
1 2 3 4 5
6 7 8 9 0

Now of course we could submit 150 jobs using LSF or we could use one job which processes the files one after another, but the most elegant way would be to submit one job for 64 cores (e.g. a whole node on Mogon I) and process the files in parallel. This is especially convenient, since we can then use the nodeshort queue which has better scheduling characteristics than short (while both show better scheduling compared to there long counterparts:

parallel_job
#!/bin/bash
 
#SBATCH --job-name=demo_gnu_parallel
#SBATCH --output=res_gnu_parallel.txt
#SBATCH --ntasks=4
#SBATCH --time=10:00
#SBATCH --mem-per-cpu=100
#SBATCH -p short
#SBATCH -A <your account>
 
# will load the most recent version of ''parallel''
module load tools/parallel
 
# Store working directory to be safe
SAVEDPWD=$(pwd)
# set jobdir
export JOBDIR=/localscratch/$SLURM_JOB_ID
 
# suppose we want to process 150 data files, we need to create them for the purpose of the example:
for ((i=0; i < 151; i++)); do
    fname="data_$(printf "%03d" $i).in"
    echo "{0..4}" >> $fname
    echo "{5..9}" >> $fname
done
 
# First, we copy the input data files and the program to the local filesystem of our node 
# (we pretend it is useful - an actual use case are programs with random I/O) on those files
cp "${SAVEDPWD}"/data_*.in $JOBDIR
 
# Change directory to jobdir
cd $JOBDIR
 
# we could set the number of threads for the program to use like this:
# export OMP_NUM_THREADS=4
# but in this case the program is not threaded
 
# -t enables verbose output to stderr
# We could also set -j $((LSB_DJOB_NUMPROC/OMP_NUM_THREADS)) to be more dynamic
# The --delay parameter should be used to distribute I/O load at the beginning of program execution by
#   introducing a delay of 1 second before starting the next task
# --progress will output the current progress of the parallel task execution
# {} will be replaced by each filename
# {#} will be replaced by the consecutive job number
# Both variants will have equal results:
#parallel -t -j 16 --delay 1 --progress "./program {/} > {/.}.out" ::: data_*.in
find . -name 'data_*.in' | parallel -t -j $SLURM_NPROCS  "wc {/} > {/.}.out"
 
# See the GNU Parallel documentation for more examples and explanation
 
# Now capture exit status code, parallel will have set it to the number of failed tasks
STATUS=$?
 
# Copy output data back to the previous working directory
cp $JOBDIR/data_*.out $SAVEDPWD/
 
exit $STATUS
$ sbatch parallel_example_script.sh

After this job has run, we should have the results/output data (in this case, it's just the output of wc, for demonstration):

$ ls data_*.out
data_001.out
data_002.out
[...]
data_149.out
data_150.out
$ cat data_001.out
2 10 20 data_001.in

Let's further assume that our program is able to work in parallel itself using OpenMP. We determined that OMP_NUM_THREADS=8 is the best amount of parallel work for one set of input data. This means we can launch 64/8=8 processes using GNU Parallel on the one node we have

parallel_job2
#!/bin/bash
#SBATCH --job-name=demo_gnu_parallel
#SBATCH --output=res_gnu_parallel.txt
#SBATCH --cpus-per-task=8
#SBATCH --time=10:00
#SBATCH --mem-per-cpu=100
#SBATCH -p short
#SBATCH -A <your account>
 
# will load the most recent version of ''parallel''
module load tools/parallel
 
# Store working directory to be safe
SAVEDPWD=$(pwd)
 
JOBDIR=/localscratch/$SLURM_JOBID
RAMDISK=$JOBDIR/ramdisk
 
export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK
 
# -t enables verbose output to stderr
# We could also set -j $((LSB_DJOB_NUMPROC/OMP_NUM_THREADS)) to be more dynamic
# The --delay parameter is used to distribute I/O load at the beginning of program execution by
#   introducing a delay of 1 second before starting the next task
# --progress will output the current progress of the parallel task execution
# {} will be replaced by each filename
# {#} will be replaced by the consecutive job number
# Both variants will have equal results:
#parallel -t -j 16 --delay 1 --progress "./program {/} > {/.}.out" ::: data_*.in
find . -name 'data_*.in' | parallel -t -j 8 --delay 1 --progress "./program {/} > {/.}.out"
# See the GNU Parallel documentation for more examples and explanation
 
# Now capture exit status code, parallel will have set it to the number of failed tasks
STATUS=$?
 
exit $STATUS

This example shows how to use user-defined functions, variables and anonymous pipes in bash. It uses bwa, a bio-informatics tool to map unknown DNA-sequences to a given reference. The reason, why it is chosen, is that it requires different kind of inputs: A reference file or directory and at least two input files. These input files may not be compressed, but, as the script shows, uncompression be means of a gzip / zcat and redirection is a working solution.

  • Note, that this example sets set -x to print command traces. This is intended merely to ease comprehension. For a script in production, this should be out-commented.
  • Also note, that information about the function is carried to the sub-shells with the export -f statement.
  • Variables which are not stated upon the call of GNU parallel can be made available to a function with additional export statements (here: $OUTPUTDIR).
  • Positional variables (here, just $1 in our example function) are given by the call through GNU parallel. This is useful for parameters which should change for every iteration, here: the input file name.

In particular: bwa can be slow on files bigger than a few GB. Hence, the proposed trade-off between threads and concurrently running processes, the load balancing, might not be optimal. (Many smaller files will probably analyzed faster with only 4 threads and 16 concurrently running processes (the -j option).)

#!/bin/bash
 
#SBATCH --job-name=bwa_demo_gnu_parallel
#SBATCH --output=res_bwa_gnu_parallel.log
#SBATCH -N 1
#SBATCH --time=300
#SBATCH -p nodeshort
#SBATCH -A <your account>
#SBATCH --gres=ramdisk:30G
 
 
# This script is written by Christian Meesters (HPC-team, ZDV, Mainz)
# 
# Please note: It is valid for Mogon I. The following restrictions apply:
# - if your fastq-files in the defined inputdirectory are big, the given
#   memory might not be sufficient. In this case restrict yourself to 
#   a data subset.
 
# in order to see the output of all commands, we set this:
set -x
 
#1 we purge all possible module to avoid a mangled setup
module purge
 
#2 we load out GNU parallel module (latest version)
module load tools/parallel
 
#3 in order perform our alignment (here: bwa sampe) and subsequent sorting we
#  load bwa and samtools
module load bio/BWA
module load bio/SAMtools
 
#4 make the return value of the last pipe command which fails the return value
set -o pipefail
 
#5 set a path the reference genome, extract its directory path
REFERENCEGENOME="reference/hg19.fasta"
REFERENCEDIR=$(dirname $REFERENCEGENOME)
 
#6 select a base directory for all input and traverse through it:
INPUTBASEDIR=./input
#6b now we gather all input files:
#  - a first fastq file (ending on _1.fastq)
#  - its mate (ending on _2.fastq)
# If your files name use a different scheme, adjust this script
FORWARD_READS=$(find -L $INPUTBASEDIR -type f -name '*_1.fastq*')
 
#7 create an output directory, here: according to bwa and samtools versions
BWA_VERSION=$(bwa |& grep Version | cut -d ' ' -f2 | cut -d '-' -f1 )
export OUTPUTDIR="bwa${BWA_VERSION}_samtools_${SLURM_JOB_ID}"
 
if [ ! -d "$OUTPUTDIR" ]; then
    mkdir -p "$OUTPUTDIR"
fi
 
#8 copy the reference to the ramdisk
NEWREFERENCEDIR=/localscratch/${SLURM_JOB_ID}/ramdisk/reference
mkdir -p $NEWREFERENCEDIR
 
for FILE in $REFERENCEDIR/*; do
  sbcast -f $FILE $NEWREFERENCEDIR/$(basename $FILE)
done
 
REFERENCEGENOME=$NEWREFERENCEDIR/$(basename $REFERENCEGENOME)
REFERENCEDIR=$NEWREFERENCEDIR
 
#9 create an alignment function with the appropriate calls for bwa and samtools
function bwa_aln {
  TEMPOUT=$(basename $1)
  # check file ending: is the file ending on gz?
  if [ "$1" == "*.gz" ]; then
        #bwa sampe $REFERENCEGENOME <( bwa aln -t 4 $REFERENCEGENOME <(zcat $1) ) \
        #                           <( bwa aln -t 4 $REFERENCEGENOME <(zcat ${1/_1/_2} ) ) \
        #                           <(zcat $1) <(zcat ${1/_1/_2}) | \
        #samtools view -Shb /dev/stdin > "$OUTPUTDIR/${TEMPOUT%_1.fastq.gz}_aligned.bam"
        bwa mem -M -t 8 $REFERENCEGENOME <(zcat $1) <(zcat ${1/_1/_2})  |
        samtools view -Shb /dev/stdin > "$OUTPUTDIR/${TEMPOUT%_1.fastq}_aligned.bam"
  else
        #bwa sampe $REFERENCEGENOME <( bwa aln -t 4 $REFERENCEGENOME $1 ) \
        #                           <( bwa aln -t 4 $REFERENCEGENOME ${1/_1/_2} ) \
        #                           $1 ${1/_1/_2} | \
 
        bwa mem -M -t 8 $REFERENCEGENOME $1 ${1/_1/_2} |
        samtools view -Shb /dev/stdin > "$OUTPUTDIR/${TEMPOUT%_1.fastq}_aligned.bam"
  fi
}
 
#9b we need to export this function, such that all subprocesses will see it (only works in bash)
export -f bwa_aln
 
# finally we start processing
#  we consider taking 4 thread for each call of bwa aln, hence 8 threads
#  and 64 / 8 is 8. This results in a little over subscription, as bwa mem
#  runs with 8 threads and samtools is run, too.
#  Note the ungrouping of output with the -u option.
parallel -v -u --env bwa_aln --no-notice -j 8 bwa_aln ::: $FORWARD_READS

We do not recommend supplying a hostlist to GNU parallel with the -S option, as GNU parallel attempts to ssh on the respective nodes (inluding the master host) and therefore looses the environment. You can script around this, but you will run into a quotation hell.

Alas, the following example does not work due to the current settings within SLURM. We strive to solve this issue. Until then, please use the example below.

Instead, we recommend a setup similar to

multi_host
#!/bin/bash
#SBATCH -J <your meaningful job name>
#SBATCH -A <your account>
#SBATCH -p nodeshort # for Mogon I
#SBATCH -p parallel  # for Mogon II
#SBATCH --nodes=3 # appropriate number of Nodes
#SBATCH -n 192    # example value for Mogon I, see below
#SBATCH -t 300
#SBATCH --cpus-per-task=8 # we assume an application which scales to 8 threads, but
                          # -c / --cpus-per-task could also be ommited (default is =1)
                          # or set to a different value.
#SBATCH -o <your logfile prefix>_%j.log
 
#adjust / overwrite those two commands to enhance readability & overview
# parameterize srun
srun="srun -N1 -n 1 -c $SLURM_CPUS_PER_TASK  --jobid $SLURM_JOBID --cpu_bind=q"
# parameterize parallel
parallel="parallel -j $SLURM_NTASKS --no-notice "
 
# your preprocessing goes here
 
# start the run with GNU parallel
$parallel $srun <command> ::: <parameter list>

The number of tasks given by -n should be the number of CPUs * the number of nodes. However, bear in mind that the a-nodes of Mogon I have 1 FPU per 2 CPU Module and the z-nodes of Mogon II have 20 CPUs, each with hyptherthreading enables. Wich number you best assume to be the number of cores is application depended and should best be determinded experimentally.

This solution is an alternate generic solution for the challenge to run a GNU-parallel job using several hosts on mogon.

multi_host2
#!/bin/bash
#SBATCH -J <your meaningful job name>
#SBATCH -A <your account>
#SBATCH -p nodeshort # for Mogon I
#SBATCH -p parallel  # for Mogon II
#SBATCH --nodes=3 # appropriate number of Nodes
#SBATCH -n 192    # example value for Mogon I, see below
#SBATCH -t 300
 
cmd=$(pwd)/cmd_file.sh
 
# write a (temporary command file)
echo '#!/bin/bash' > $cmd
 
# set up the environment within that file
echo 'module load <your desired modules' >> $cmd
printf "$your_defined_command <some args>" >> $cmd
# arguments which require string manipulation need different quoting, e.g.:
echo '--input ./$1 --output ./output/$(basename ${1%input_suffix}out)' >> $cmd
chmod +x $cmd
 
HOSTLIST=$(scontrol show hostname $SLURM_JOB_NODELIST | paste -d, -s )
 
# start the parallel work on the defined hosts
parallel  --workdir $PWD --sshdelay 0.2 -S $HOSTLIST -j <number of parallel tasks per host>  $cmd :::<parameter list>       

SLURM multiprog for uneven arrays

The SLURM multiprog option in srun essentially displays a master-slave setup. You need it to run within a SLRUM job allocation and trigger srun with the --multi-prog option and appropriate multiprog file:

master_slave_simple.sh
#!/bin/bash
#
#SBATCH --job-name=test_ms
#SBATCH --output=res_ms.txt
# parameters of this snippet, choose sensible values for your setup
#SBATCH --ntasks=4
#SBATCH --time=10:00
#SBATCH --mem-per-cpu=100
 
# for the purpose of this course
#SBATCH -A <your account>
#SBATCH -p short
 
srun <other parameters> --multi-prog multi.conf

Then, of course the multi.conf file has to exist:

multi.conf
0      echo     'I am the Master'
1-3    bash -c 'printenv SLURM_PROCID'

Indeed, as the naming suggests, you can use such setup to emulate a master-slave environment. But then the processes have to care themselves about there communication (sockets, regular files, etc.). And the most cumbersome aspect is: You have to maintain two files at all times, whenever the setup has to be changed, and all parameters have to match.

The configuration file contains three fields, separated by blanks. These fields are :

  • Task number
  • Executable File
  • Argument

Parameters available :

  • %t - The task number of the responsible task
  • %o - The task offset (task's relative position in the task range).

The ZDV-taskfarm Script an alternative to multiprog

The script is hosted on github, forked from Paddy Doyle, Trinity College, Dublin and adapted for Mogon (I and II).

The slurm multi-prog setup can be difficult for some scenarios:

  • only one executable can be specified per task (e.g. no chain of commands or shell loops are possible, such as cd dir01; ./my_exec)
  • limitation on the maximum number of characters per task description (256)
  • building the multi-prog file can be onerous, if you do not have the luxury of using the '%t' tokens in your commands or arguments
  • the number of commands must match exactly the number of slurm tasks (-n), which means updating two files if you wish to add or remove tasks

Slurm Job Arrays are a better option to multi-prog, unless using the parallel (Mogon II) or node* (Mogon I) partitions, with scalable software, anyway.

The taskfarm script makes using multi-prog setups easy. Please only use it, if your tasks have +/- the same run time or else huge parts of the reserved nodes can be left idle.

For a full listing of the command line interface you can load the module and ask the script itself for help:

$ module load tools/staskfarm
$ staskfarm -h
taskfarm_file
#!/bin/bash
 
#SBATCH -J taskfarm_example
#SABTCH -o taskfarm_example_%j.out
#SBATCH -N2 # in this example we take 2 nodes
#SBATCH -n 128 # optional argument - the optimal setting (or ommitance) has to be tried on a case basis
#SBATCH -A <your account>
#SBATCH -p nodeshort
 
# will load the most recent module version of the taskfarm
module load tools/staskfarm
 
# - suppose we have a program which requires 2 intputs:
#    'input_<number>_R1.fastq' and 'input_<number>_R2.fastq'
# - assume further we have 302 such files
# - and we want to work on them in a round robin manner
 
# 1st we "produce" such dummy files:
for ((i=0; i < 303; i++)); do
   touch "input_${i}_R1.fastq" 
   touch "input_${i}_R2.fastq" 
done
 
# 3rd, we specify our input command.
#      Instead of a 'real' application we drop in 'echo'.
#      And we use pattern expansion to retrieve the 2nd file name,
#      as we cannot (always) loop over several expressions.
echo '#!/bin/bash' > cmd_file.sh
echo 'echo "working on node $(hostname) on files $1 and ${1%%_*}_R2.fastq"' >> cmd_file.sh
chmod +x cmd_file.sh
cmd=$(pwd)/cmd_file.sh
 
# 4th, start the taskfarm:
staskfarm $cmd *_R1.fastq
 
# finally, we need to clean up our mess:
rm *fastq cmd_file.sh

As stated, the most sensible use case for the taskfarm are application with +/- equal run times for the inputs. When screening parameters in a simulation, is is likely that run times greatly depend on the parameters. Therefore, it might be better to consider using GNU parallel or other parallelisation schemes.

taskfarm_file2
#!/bin/bash
 
#SBATCH -J taskfarm_example
#SABTCH -o taskfarm_example_%j.out
#SBATCH -N1 # in this example we take <1 node
#SBATCH -n 40
#SBATCH --cpus-per-task=2
#SBATCH -A <your account>
#SBATCH --time 10
#SBATCH -p short
 
# will load the most recent module version of the taskfarm
module load tools/staskfarm
 
# - suppose we have a program which requires 2 intputs:
#    all natural numbers for arg1 between 1 and 5 and for arg2 just 0 and 1
# - assume further we want to run a permutation test
 
# 1st we "produce" such dummy files:
permutations=$(echo {1..5},{0..1})
 
# 3rd, we specify our input command.
#      Instead of a 'real' application we drop in 'echo'.
echo '#!/bin/bash' > cmd_file.sh
echo 'echo "working on node $(hostname) on args $1 and with $OMP_NUM_THREADS threads' >> cmd_file.sh
chmod +x cmd_file.sh
cmd=$(pwd)/cmd_file.sh
 
# 4th, start the taskfarm:
#      NOTE: We start each application with 2 threads (-t 2)
# the '--parameters' flag, will skip the test for input files (as parameters aren't files)
staskfarm -t 2 --parameters $cmd $permutations
 
# finally, we need to clean up our mess:
rm cmd_file.sh
  • node_local_scheduling.1510662645.txt.gz
  • Last modified: 2017/11/14 13:30
  • by meesters