Search Docs by Keyword

Table of Contents

Submitting Large Numbers of Jobs to the FASRC cluster



Often times one will need to submit a large number of jobs to the cluster for various reasons — to get high throughput for a big analysis, to vary parameters with one analysis, etc. This document aims to help you become more efficient and to help you take advantage of shell and SLURM resources. This will improve your work and help others on the cluster!

A word of warning, though. Submitting large numbers of jobs to the cluster can have disastrous consequences if not done correctly, as one can overload the scheduler, bringing the cluster to a grinding halt. Please use these best practices (as outlined in our Customs and Responsibilities document):

  • Please pause 0.5 to 1 seconds between each sbatch submit
  • Please ensure that each job takes at least 5 to 10 minutes to run, in order to allow the scheduler to process all the work needed to set up, run, and break down scheduled jobs.
  • Please do not use email notifications except for errors. Jobs that overwhelm the email queue will be killed.

We do also recommend that people investigate the use of gnu_parallel or TACC’s launcher, which is installed on the cluster. We’ll include more documentation on these programs in the near future.
We make a few assumptions in this document:

  • your filenames do not contain spaces. If so, many of these commands may fail.
  • you are familiar with basic UNIX concepts
  • you understand the partitions and filesystems that are available to the compute nodes.

Using shell commands

It is relatively easy to use the shell to submit jobs to SLURM:

# loop over all FASTA files in the directory, print the filename
# (so we have some visual progress indicator), then submit the
# gzip jobs to SLURM
for FILE in *.fasta; do
echo ${FILE}
sbatch -p serial_requeue -t 10 --mem=200 --wrap="gzip ${FILE}"
sleep 1 # pause to be kind to the scheduler

These commands do exactly what the comments suggest, sending gzip jobs to compute nodes on the cluster in a high-throughput manner. A couple of things to note:

  • We send jobs to serial_requeue to ensure that the cluster is efficiently used for small (< 24 hr) jobs.
  • Often times you’ll see shell variables written like $FILE. $FILE and ${FILE} are exactly the same, but ${FILE} is safer to use, especially when embedded in a string of characters. For example myfile_$DATE_normal.txt is not the same as myfile_${DATE}_normal.txt, as in the former, the shell will look for the variable $DATE_normal in the former example and for $DATE in the latter one.
  • Using the --wrap= option for SLURM is generally not advised except for very limited cases as you will not have your normal bash environment, only a bare sh shell. In this example it allows you to avoid writing large numbers of batch files for very simple commands. But this is really only useful for one-line commands as, again, wrap will not execute your bash shell environment (and .bashrc, etc.), but a base sh shell only.

This script is great for a small number of files that you wish to do some trivial operations on (e.g. 3 – 20ish?). Since it is asking SLURM to process each file individually, SLURM will have to track each job/file. OK in small numbers; not so good for large numbers. Also, if you ask SLURM to handle 20 jobs as once, and all those jobs are working on the same directory, you might slow down disk access, esp. if your files are located on a lab disk share. In this case it we advise that you work on one of the scratch filesystems, disks and directories that have been optimized for high-throughput disk access. Please see our guides Where to Store Your Files and the cluster Storage for more information. In addition, since it takes SLURM a minimum of several minutes to calculate where and when to run your job, it is best if your jobs/files take at least a minimum of 5 minutes to run.
For more complex jobs, you may need to use a SLURM batch file. For example, to run the bioinformatics program tophat with a bunch of FASTQ files, create a SLURM batch file tophat_manyfiles.sbatch:

# tophat_manyfiles.sbatch
#SBATCH -p serial_requeue # Partition
#SBATCH -c 1 # one CPU
#SBATCH -t 0-2:00 # Running time of 2 hours
#SBATCH --mem 4000 # Memory request
mkdir ${1}_out
cd ${1}_out
tophat /n/informatics/databases/Mus_musculus/UCSC/mm10/Sequence/BowtieIndex \

And you submit your jobs with the command:

# loop over all FASTQ files in the directory
# print the filename (so we have some visual progress indicator)
# then submit the tophat jobs to SLURM
for FILE in *.fq; do
echo ${FILE}
sbatch -o ${FILE}.stdout.txt -e ${FILE}.stderr.txt tophat_manyfiles.sbatch ${FILE}
sleep 1 # pause to be kind to the scheduler

A few explanations:

  • In the bash loop, each FASTQ filename will be substituted for ${FILE}.
  • Inside the SLURM batch file, the passed filename will be substituted for ${1}. In bash, ${1} is the 1st parameter passed to the script file.
  • Normally, the STDOUT and STDERR options (-o and -e) are placed in the SLURM batch file. Unfortunately, since the #SBATCH lines start with # (the hash symbol), bash sees these as comment lines and so it won’t do any substitution. To compensate, we put the -o and -e parameters directly on the sbatch call in the shell loop.
  • In our SLURM batch file, we create a directory named similarly to the input FASTQ file, move into that directory, then run tophat. Nota bene: As we discussed previously, ${1}_out is not the same as $1_out!

Using job arrays

Although the previous example is an effective method, it is not very efficient for SLURM, since it has to keep track of and schedule each individual job. A better way to do this is to use job arrays. This is documented at this RC job arrays page and in more detail at the official SLURM job arrays page. This requires that your files be named in sequential order. Briefly, create your SLURM batch file tophat_arrays.sbatch:

# tophat_arrays.sbatch
#SBATCH -J tophat # A single job name for the array
#SBATCH -p serial_requeue # Partition
#SBATCH -c 1 # one core
#SBATCH -t 0-2:00 # Running time of 2 hours
#SBATCH --mem 4000 # Memory request of 4 GB
#SBATCH -o tophat_%A_%a.out # Standard output
#SBATCH -e tophat_%A_%a.err # Standard error
#SBATCH --array=1-30  # Size of the array
mkdir trans${SLURM_ARRAY_TASK_ID}_out
cd trans${SLURM_ARRAY_TASK_ID}_out
tophat /n/informatics/databases/Mus_musculus/UCSC/mm10/Sequence/BowtieIndex \

And we submit one, grouped set of jobs to SLURM with:

sbatch tophat_arrays.sbatch
If you want to test for small set
sbatch --array=1-5 tophat_arrays.sbatch 

Much simpler, huh? Explanations:

  • %A in the #SBATCH line becomes the job ID
  • %a in the #SBATCH line becomes the array index
  • ${SLURM_ARRAY_TASK_ID} is a shell variable that is set when the job runs, and it is substituted into the parameter to generate the proper filename which is passed on to tophat

Using job arrays with non-sequentially named files

Well, job arrays are great if your files are named sequentially (e.g. file1.fq, file2.fq, etc). But what if they’re not? One can use both shell arrays + SLURM arrays to solve this. A little tricky, but manageable…
You’ll read in your list of files into a bash array (e.g. FILES), and then work with one item in this list (array) inside your SLURM script. Your script will know which item it is using the shell variable ${SLURM_ARRAY_TASK_ID} we used last time. Briefly, create your SLURM batch file tophat_double_array.sbatch:

# tophat_double_array.sbatch
#SBATCH -J tophat # A single job name for the array
#SBATCH -p serial_requeue # best partition for single core small jobs
#SBATCH -c 1 # one core
#SBATCH -t 0-2:00 # Running time of 2 hours
#SBATCH --mem 4000 # Memory request of 4 GB
#SBATCH -o tophat_%A_%a.out # Standard output
#SBATCH -e tophat_%A_%a.err # Standard error
# grab out filename from the array exported from our 'parent' shell
# make & move into new directory, and run!
mkdir ${FILENAME}_out
cd ${FILENAME}_out
tophat /n/informatics/databases/Mus_musculus/UCSC/mm10/Sequence/BowtieIndex \

Now, we grab all the appropriate files and submit them en-batch with an array:

# grab the files, and export it so the 'child' sbatch jobs can access it
export FILES=($(ls -1 *.fq))
# get size of array
# now subtract 1 as we have to use zero-based indexing (first cell is 0)
# now submit to SLURM
if [ $ZBNUMFASTQ -ge 0 ]; then
sbatch --array=0-$ZBNUMFASTQ tophat_double_array.sbatch

Not so hard, huh? Some explanations:

  • Bash arrays are indexed using [] characters at the end of the array variable name wrapped with {}
  • The strange =($(ls -1 *.fq)) syntax combines executing a command via $(unix command) with array assignment =(array items)
  • One can get the length of a bash array by indexing the array with [@] and prefixing the variable name with #
  • Since bash arrays start with 0 (e.g. $FILES[0]), we need to adjust our array to do so as well. So we do bash arithmetic using the $((math expression)) syntax
  • Finally, so that we don’t throw bogus jobs at the scheduler, we use a bash conditional if [ test ]; then ... fi to ensure that our number of items is greater than (-gt) zero.

Varying parameters to batch jobs

(If you’re still with us, great!) So, a routine exercise for R or MATLAB scripts is to run an analysis varying the input parameters. Of course, it’d be great if we can batch these out to the cluster all at once! Create your SLURM batch file vary_params.sbatch:

# vary_params.sbatch
#SBATCH -p serial_requeue # Partition
#SBATCH -c 1 # one core
#SBATCH -t 0-0:30 # Running time of 30 minutes
#SBATCH --mem 1000 # Memory request ${PARAM1} ${SAMPLE_SIZE}

And at the shell,

for PARAM1 in $(seq 1 5); do
for SAMPLE_SIZE in 25 50 75 100; do
echo "${PARAM1}, ${SAMPLE_SIZE}"
sbatch -o out_p${PARAM1}_s${SAMPLE_SIZE}.stdout.txt \
-e out_p${PARAM1}_s${SAMPLE_SIZE}.stdout.txt \
--job-name=my_analysis_p${PARAM1} \
sleep 1 # pause to be kind to the scheduler


  • The bit of code $(seq 1 5) runs the seq command in a subprocess that will generate the sequence 1..5. We could have easily written 1 2 3 4 5. But now you’ve learned a new trick.
  • We create a nested loop with SAMPLE_SIZE.
  • The export command is needed to ensure that the shell makes the environment variables PARAM1 and SAMPLE_SIZE available to any sub-shells (your sbatch script).
  • In our sbatch command, to keep the command easier to read, we terminate each line of the same command with the backslash character . No other character of any kind can follow the backslash on the line, or else the backslash will be ignored. Bummer!
  • So that we can keep all our output in order and since shell variables are not expanded on the #SBATCH line, we specify the STDOUT (-o) and STDERR (-e) on the sbatch line.
  • To track our jobs more easily, we’re going to group the job names by PARAM1. We can do this by using the --job-name= parameter on the sbatch line. Now, all the sample size jobs with the same PARAM1 value will have the same job name, and we can control them as a group (e.g. in case we need to kill the jobs).

For further reading


© The President and Fellows of Harvard College
Except where otherwise noted, this content is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license.