Array jobs: embarassingly parallel execution

Videos

Videos of this topic may be available from one of our kickstart course playlists: 2023, 2022 Summer, 2022 February, 2021 Summer, 2021 February.

Abstract

  • Arrays allow you to submit jobs and it runs many times with the same Slurm parameters.

  • Submit with the --array= Slurm argument, give array indexes like --array=1-10,12-15.

  • The $SLURM_ARRAY_TASK_ID environment variable tells a job which array index it is.

  • There are different templates to use below, which you can adapt to your task.

  • If you aren’t fully sure of how to scale up, contact us Research Software Engineers early.

More often than not, scientific problems involve running a single program again and again with different datasets or parameters.

When there is no dependency or communication among the individual program runs, these individual runs can be run in parallel on separate Slurm jobs. This kind of parallelism is called embarassingly parallel.

Slurm has a structure called job array, which enables users to easily submit and run several instances of the same Slurm script independently in the queue.

Schematic of cluster with current discussion points highlighted; see caption or rest of lesson.

Array jobs let you control a large amount of the cluster. In Parallel computing: different methods explained, we will see another way.

Introduction

Array jobs allow you to parallelize your computations. They are used when you need to run the same job many times with only slight changes among the jobs. For example, you need to run 1000 jobs each with a different seed value for the random number generator. Or perhaps you need to apply the same computation to a collection of data sets. These can be done by submitting a single array job.

A Slurm job array is a collection of jobs that are to be executed with identical parameters. This means that there is one single batch script that is to be run as many times as indicated by the --array directive, e.g.:

#SBATCH --array=0-4

creates an array of 5 jobs (tasks) with index values 0, 1, 2, 3, 4.

The array tasks are copies of the submitted batch script that are automatically submitted to Slurm. Slurm provides a unique environment variable SLURM_ARRAY_TASK_ID to each task which could be used for handling input/output files to each task.

../../../_images/parallel-array.svg

--array via the command line

You can also pass the --array option as a command-line argument to sbatch. This can be great for controlling things without editing the script file.

Important

When running array job you’re basically running identical copies of a single job. Thus it is increasingly important to know how your code behaves with respect to the file system:

  • Does it use libraries/environment stored in the work directory?

  • How much input data does it need?

  • How much output data does the job create?

For example, running an array job with hundreds of workers that uses a Python environment stored in the work disk can inadvertently cause a lot of filesystem load as there will be hundreds of thousands of file calls.

If you’re unsure how your job will behave, ask us Research Software Engineers for help for help.

Your first array job

Let’s see a job array in action. Lets create a file called array_example.sh and write it as follows.

#!/bin/bash
#SBATCH --time=00:15:00
#SBATCH --mem=200M
#SBATCH --output=array_example_%A_%a.out
#SBATCH --array=0-15

# You may put the commands below:

# Job step
srun echo "I am array task number" $SLURM_ARRAY_TASK_ID

Submitting the job script to Slurm with sbatch array_example.sh, you will get the message:

Submitted batch job 60997836

The job id in the message is that of the primary array job. This is common for all of the jobs in the array. In addition, each individual job is given an array task id.

As now we’re submitting multiple jobs simultaneously, each job needs an individual output file or the outputs will overwrite each other. By default, Slurm will write the outputs to files named slurm-${SLURM_ARRAY_JOB_ID}_${SLURM_ARRAY_TASK_ID}.out. This can be overwritten using the --output=FILENAME-parameter, when you can use wildcard %A for the job id and %a for the array task id.

Once the jobs are completed, the output files will be created in your work directory, with the help %u to determine your user name:

$ ls
array_example_60997836_0.out   array_example_60997836_12.out  array_example_60997836_15.out  array_example_60997836_3.out  array_example_60997836_6.out  array_example_60997836_9.out
array_example_60997836_10.out  array_example_60997836_13.out  array_example_60997836_1.out   array_example_60997836_4.out  array_example_60997836_7.out  array_example.sh
array_example_60997836_11.out  array_example_60997836_14.out  array_example_60997836_2.out   array_example_60997836_5.out  array_example_60997836_8.out

You can cat one of the files to see the output of each task:

$ cat array_example_60997836_11.out
I am array task number 11

Important

The array indices do not need to be sequential. For example, if after running an array job you find out that tasks 2 and 5 failed, you can relaunch just those jobs with --array=2,5.

You can even simply pass the --array option as a command-line argument to sbatch.

More examples

The following examples give you an idea on how to use job arrays for different use cases and how to utilize the $SLURM_ARRAY_TASK_ID environment variable. In general,

  • You need some map of numbers to configuration. This might be files on the filesystem, a hardcoded mapping in your code, or some configuration file.

  • You generally want the mapping to not get lost. Be careful about running some jobs, changing the mapping, and running more: you might end up with a mess!

Reading input files

In many cases, you would like to process several data files. That is, pass different input files to your code to be processed. This can be achieved by using $SLURM_ARRAY_TASK_ID environment variable.

In the example below, the array job gives the program different input files, based on the value of the $SLURM_ARRAY_TASK_ID:

#!/bin/bash
#SBATCH --time=01:00:00
#SBATCH --mem=1G
#SBATCH --array=0-29

# Each array task runs the same program, but with a different input file.
srun ./my_application -input input_data_${SLURM_ARRAY_TASK_ID}

Hardcoding arguments in the batch script

One way to pass arguments to your code is by hardcoding them in the batch script you want to submit to Slurm.

Assume you would like to run the pi estimation code for 5 different seed values, each for 2.5 million iterations. You could assign a seed value to each task in you job array and save each output to a file. Having calculated all estimations, you could take the average of all the pi values to arrive at a more accurate estimate. An example of such a batch script pi_array_hardcoded.sh is as follows.

#!/bin/bash
#SBATCH --time=01:00:00
#SBATCH --mem=500M
#SBATCH --job-name=pi-array-hardcoded
#SBATCH --output=pi-array-hardcoded_%a.out
#SBATCH --array=0-4

case $SLURM_ARRAY_TASK_ID in
   0)  SEED=123 ;;
   1)  SEED=38  ;;
   2)  SEED=22  ;;
   3)  SEED=60  ;;
   4)  SEED=432 ;;
esac

srun python slurm/pi.py 2500000 --seed=$SEED > pi_$SEED.json

Save the script and submit it to Slurm:

$ sbatch pi_array_hardcoded.sh
Submitted batch job 60997871

Once finished, 5 Slurm output files and 5 application output files will be created in your current directory each containing the pi estimation; total number of iterations (sum of iteration per task); and total number of successes):

$ cat pi_22.json
{"successes": 1963163, "pi_estimate": 3.1410608, "iterations": 2500000}

Reading parameters from one file

Another way to pass arguments to your code via script is to save the arguments to a file and have your script read the arguments from it.

Drawing on the previous example, let’s assume you now want to run pi.py with different iterations. You can create a file, say iterations.txt and have all the values written to it, e.g.:

$ cat iterations.txt
100
1000
50000
1000000

You can modify the previous script to have it read the iterations.txt one line at a time and pass it on to pi.py. Here, sed is used to get each line. Alternatively you can use any other command-line utility, e.g. awk. Do not worry if you don’t know how sed works - Google search and man sed always help. Also note that the line numbers start at 1, not 0.

The script pi_array_parameter.sh looks like this:

#!/bin/bash
#SBATCH --time=01:00:00
#SBATCH --mem=500M
#SBATCH --job-name=pi-array-parameter
#SBATCH --output=pi-array-parameter_%a.out
#SBATCH --array=1-4

n=$SLURM_ARRAY_TASK_ID
iteration=`sed -n "${n} p" iterations.txt`      # Get n-th line (1-indexed) of the file
srun python slurm/pi.py ${iteration} > pi_iter_${iteration}.json

You can additionally do this procedure in a more complex way, e.g. read in multiple arguments from a csv file, etc.

(Advanced) Two-dimensional array scanning

What if you wanted an array job that scanned a 2D array of points? Well, you can map 1D to 2D via the following pseudo-code: x = TASK_ID // N (floor division) and y = TASK_ID % N (modulo operation). Then map these numbers into your grid. This can be done in bash, but at this point you’d want to start thinking about passing the SLURM_ARRAY_TASK_ID variable into your code itself for this processing.

(Advanced) Grouping runs together in bigger chunks

If you have lots of jobs that are short (a few minutes), using array jobs may induce too much overhead in scheduling and you will create huge number of output files. In these kinds of cases you might want to combine multiple program runs into a single array job.

Important

A good target time for the array jobs would be approximately 30 minutes, so please try to combine your tasks so that each job would at least take this long.

Easy workaround for this is to create a for-loop in your Slurm script. For example, if you want to run the pi script with 50 different seed values you could run them in chunks of 10 and run a total of 5 array jobs. This reduces the amount of array jobs we need by a factor of 10!

This method demands more knowledge of shell scripting, but the end result is a fairly simple Slurm script pi_array_parameter.sh that does what we need.

#!/bin/bash
#SBATCH --time=01:00:00
#SBATCH --mem=500M
#SBATCH --job-name=pi-array-grouped
#SBATCH --output=pi-array-grouped_%a.out
#SBATCH --array=1-4

# Lets create a new folder for our output files
mkdir -p json_files

CHUNKSIZE=10
n=$SLURM_ARRAY_TASK_ID
indexes=`seq $((n*CHUNKSIZE)) $(((n + 1)*CHUNKSIZE - 1))`

for i in $indexes
do
   srun python slurm/pi.py 1500000 --seed=$i > json_files/pi_$i.json
done

Exercises

The scripts you need for the following exercises can be found in our hpc-examples, which we discussed in Using the cluster from a shell. You can clone the repository by running git clone https://github.com/AaltoSciComp/hpc-examples.git. Doing this creates you a local copy of the repository in your current working directory. This repository will be used for most of the tutorial exercises.

Array-1: Array jobs and different random seeds

Create a job array that uses the slurm/pi.py to calculate a combination of different iterations and seed values and save them all to different files. Keep the standard output (#SBATCH --output=FILE) separate from the standard error (#SBATCH --error=FILE).

Array-2: Combine the outputs of the previous exercise.

You find an pi-aggregate.py program in hpc-examples. Run this and give all the output files as arguments. It will combine all the statistics and give a more accurate value of \(\pi\).

Array-3: Reflect on array jobs in your work

Think about your typical work. How could you split your stuff into trivial pieces that can be run with array jobs? When can you make individual jobs smaller, but run more of them as array jobs?

(Advanced) Array-4: Array jobs with advanced index selector

Make a job array which runs every other index, e.g. the array can be indexed as 1, 3, 5… (the sbatch manual page can be of help)

Array-5: Array job with varying memory requirements.

Make an array job that runs slurm/memory-use.py with five different values of memory (50M, 100M, 500M, 1000M, 5000M) using one of the techniques above - this is the memory that the memory-use script requests, not the is requested from Slurm. Request 250M of memory for the array job. See if some of the jobs fail.

Is this a proper use of array jobs?

See also

What’s next?

The next tutorial is about shared memory parallelism.