Containers¶
The login, interactive and compute nodes have Singularity CE 4.4, with no module to load. You pull a container image on a login node and run it in a job.
-
Pull or copy the image on a login node, into a directory under
/scratchor/data. Pulling downloads and converts the image, which takes minutes and up to several gigabytes: -
Run it in a job. The container sees your home directory and the current directory. Bind other directories with
--bind:samtools.job#$ -S /bin/sh #$ -N samtools -cwd -j y -o samtools.log # echo + `date` job $JOB_NAME started in $QUEUE with jobID=$JOB_ID on $HOSTNAME singularity exec --bind /scratch /scratch/genomics/USERNAME/images/samtools_1.19.2--h50ea8bc_0.sif samtools view -c reads.bam echo = `date` job $JOB_NAME done
A container that uses a GPU needs --nv on the singularity command and a GPU request on the job. ~hpc/examples/containers has working examples (lolcow.job, test1.job, test2.job) with their logs. The Singularity CE user guide at https://docs.sylabs.io/guides/latest/user-guide/ covers the command itself.
Further reading¶
- Examples for the tested container job under
~hpc/examples/containers - Singularity documentation from the project
Last updated 2026-10-03