GNU Parallel

GNU parallel is a shell tool for executing many small independent tasks on the multi-core platform (compute nodes). A job can be a single command or a small script that has to be run for each of the lines in the input. GNU parallel can then split the input and pipe it into commands in parallel.

Adding GNU Parallel to Environment

To use GNU parallel on the command line,

module load parallel

GNU Parallel Syntax

• Reading commands to be run in parallel from an input file

    parallel [OPTIONS] < CMDFILE
  

• Reading command arguments on the command line

    parallel [OPTIONS] COMMAND [ARGUMENTS] ::: ARGLIST
  

• Reading command arguments from an input file

    parallel [OPTIONS] COMMAND [ARGUMENTS] :::: ARGFILE
  

​https://www.gnu.org/software/parallel/parallel_tutorial.html

Many Task Computing

For example, you have many scripts to run,

[fuji@cypress1 JobArray2]$ ls
hello2.py    script01.sh  script03.sh  script05.sh  script07.sh  script09.sh   slurmscript2
script00.sh  script02.sh  script04.sh  script06.sh  script08.sh  slurmscript1

To run all distributed over 4 cores,

[fuji@cypress1 JobArray2]$ parallel -j 4 sh ::: script??.sh

Run Gnu Parallel in Slurm script

Single Node

The job script below requests 1 node 20 cores (one whole node). Assuming that each task is multi-thread, and OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK determines the number of threads. "-j $SLURM_NTASKS" determines the number of concurrently running tasks.

#!/bin/bash
#SBATCH --partition=defq	# Partition
#SBATCH --qos=normal		# Quality of Service
#SBATCH --job-name=GNU_Paralle 	# Job Name
#SBATCH --time=00:10:00		# WallTime
#SBATCH --nodes=1 		# Number of Nodes
#SBATCH --ntasks-per-node=5 	# Number of tasks 
#SBATCH --cpus-per-task=4 	# Number of processors per task OpenMP threads()
#SBATCH --gres=mic:0  		# Number of Co-Processors

module load parallel

export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK

parallel --record-env

parallel --joblog log \
	-j $SLURM_NTASKS \
	--workdir $SLURM_SUBMIT_DIR \
	--env OMP_NUM_THREADS \
	sh ./run_hostname.sh {} ::: `seq 1 100`

The example task script is:

#!/bin/bash
hostname
echo $1
sleep 1

Multiple Nodes

The job script below requests 4 nodes 20 cores for each. Assuming that each task is multi-thread, and OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK determines the number of threads. "-j $TASKS_PER_NODE" determines the number of concurrently running tasks per node, which is defined by TASKS_PER_NODE=echo $SLURM_NTASKS / $SLURM_NNODES | bc. scontrol show hostname $SLURM_NODELIST > machinefile makes a list of nodes and send it to GNU parallel by "—slf $MACHINEFILE".

#!/bin/bash
#SBATCH --partition=defq	# Partition
#SBATCH --qos=normal		# Quality of Service
#SBATCH --job-name=GNU_Paralle 	# Job Name
#SBATCH --time=00:10:00		# WallTime
#SBATCH --nodes=4 		# Number of Nodes
#SBATCH --ntasks-per-node=5 	# Number of tasks 
#SBATCH --cpus-per-task=4 	# Number of processors per task OpenMP threads()
#SBATCH --gres=mic:0  		# Number of Co-Processors

module load parallel

MACHINEFILE="machinefile"
scontrol show hostname $SLURM_NODELIST > machinefile
cat $MACHINEFILE
export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK
TASKS_PER_NODE=`echo $SLURM_NTASKS / $SLURM_NNODES | bc`
echo "TASKS_PER_NODE=" $TASKS_PER_NODE

parallel --record-env

parallel --joblog log \
	-j $TASKS_PER_NODE \
	--slf $MACHINEFILE \
	--workdir $SLURM_SUBMIT_DIR \
	--sshdelay 0.1 \
	--env OMP_NUM_THREADS \
	sh ./run_hostname.sh {} ::: `seq 1 100`

echo "took $SECONDS sec"