marcodelapierre / toy-gpu-nf Goto Github PK

Toy pipeline to test Nextflow with Nvidia GPUs on Topaz at Pawsey

The pipeline requires Nextflow to run, with some special requirements:

• DSL2 syntax is used, so that Nextflow version 20.07.1 or higher is required;
• Slurm multi-cluster support is needed, so that an hacked version of Nextflow is required; a script to build and install one is provided.

Here are the highlights of the key requirements to run a CPU+GPU pipeline with containers on Pawsey multiple Slurm clusters:

• the pipeline is assumed to be run from Zeus;
• hacked version of Nextflow with Slurm multi-cluster support; in this example, slurm_topaz executor is used to submit the GPU jobs to Topaz;
• process options for GPU tasks: executor = 'slurm_topaz', queue = 'gpuq', clusterOptions += " --gpus-per-node=1";
• Singularity options:
  • runOptions = "-B /group,/scratch --nv";
  • //envWhitelist = 'SINGULARITY_BINDPATH, SINGULARITYENV_LD_LIBRARY_PATH' commented out;
  • --nv is fine for non-GPU tasks, it will just print a warning in the output;
  • adding -B /group,/scratch here, because with multi-cluster submission the Singularity variables cannot be used as they differ among clusters.

Key files provided:

• main.nf
• nextflow.config, valid for test runs on Pawsey (uses debug queues)
• job.sh template Slurm script for Zeus at Pawsey
• extra/ files:
  • install-nextflow-hack-slurm-multi-cluster.sh, a script to build and install hacked multi-cluster Nextflow
  • log.sh, a helper script to get enhanced logging for completed pipelines
  • Dockerfile used to build the toy container in use
• src/: source files for the toy CPU and GPU binaries in use
• bin_/: toy binaries (directory has broken name, to force using binaries from the container)
• test-outputs/: reference test outputs (currently the pipeline only outputs the equivalent of out-a-cpu and out-a-gpu)