genomedk-tutorial

Introduction

We have all been in the same situation, where we quickly want to sort a bam file, or maybe just want to check a new tool published. Maybe we check the load on the frontend of genomedk. It does not look to be used that much; the coast is clear, and we run our software directly on the frontend, which ends up taking 8 CPUs, and suddenly it is slow for everyone.

This leads to annoyed users where a simple ls command is slow, and in absolute worst case scenario it can be dangerous. Maybe a clinical interpreter is looking at some data for a patient that needs an urgent answer. But now it is not possible, due to the heavy load on the frontend.

Is there a better way to run your jobs? How should you run a single line of code? What if it is a complex workflow? Below are listed different ways to to run software on genomedk.

Below the examples will refer to an account. This can for example be MomaDiagnostics, Improve, etc. Use the one appropriate for your project and affiliation. It is possible to read more here.

srun

In the case where you want to test how a software works, it can easily be solved with the srun command

  srun --pty --account=<my-account> /bin/bash

This will allocate 4 hours with 1GB memory on 1 CPU. If you need more resources, then do

  srun --pty --mem=16g --cpus-per-task=4 \
       --time=120 --account=<my-account> /bin/bash

Here are allocated 16GB memory, 4 CPUs, and 120 minutes. For more informaton, just run

  srun --help

to see all available options.

It might take a few moments before you enter a node depending on the resources you request, and the current load on the cluster.

srun is a great tool for exploring a new tool without putting a load on the frontend.

sbatch

While srun is good for a more interactive workflow, it is probably not what you want once you are settled on a standard way to run your new tool. This is where sbatch enters. You can make a bash script like

  #!/bin/bash
  #SBATCH --account <my-account>
  #SBATCH --mem 4g
  #SBATCH --time 12:00:00
  #SBATCH -c 4

  echo This is my first script

and simple run like

  sbatch example.sh

This will give you a jobid, which you can inspect with the jobinfo command. You can also run sbatch more directly without creating a script like

  sbatch --account <my-account> --mem 4g --time 12:00:00 -c 4 \
      echo This is my first script

However, creating a script is preferred so you are able to document in the future how you obtained the results.

gwf

gwf can be installed with conda and are great for more complex workflows. Here output from one job can be used as input for another job, and so forth.

An example could be

1. Trim fastq files
2. Map trimmed fastq files
3. Call SNVs on resulting bam file
4. Call CNVs on resulting bam file
5. ... Here item 3 and 4 can be run in parallel, since there are no dependencies between them.

Real example

It is easiest to find some examples on for example the github page of gwf, here is a good example. Otherwise you can find people at MOMA that often use gwf, for example the clinical bioinformaticians.

For running a gwf workflow you need a file called workflow.py and then run

  gwf run

To see the status of all the jobs in your current workflow, then run

  gwf status

or if you are only interested in jobs running and shouldrun, then

  gwf status -s running -s shouldrun

snakemake

Another tool for running complex workflows is snakemake. While many people at MOMA uses gwf, it is a tool probably only used on genomedk, while snakemake has a large international community.