Snakemake workflow: rna-seq-star-deseq2 (using pcluster-slurm executor)

This is my adoption of the original forked repo's rna seq start deseq2 worflow, but to use AWS Parallel Cluster, via the pcluser slurm snakemake executor.

This workflow performs a differential gene expression analysis with STAR and Deseq2.

Usage

Have an AWS Parallel Cluster ( using slurm as the scheduler ) Running.

From the headnode, proceed to the following steps.

Prereq : Conda

• Have conda activated.

Clone Repo (it includes sample data)

git clone git@github.com:Daylily-Informatics/rna-seq-star-deseq2.git
cd rna-seq-star-deseq2

Build The Snakemake (v8.*) Conda Env

conda create -n snakemake -c conda-forge  snakemake==8.24 snakedeploy tabulate yaml
conda activate snakemake
pip install snakemake-executor-plugin-pcluster-slurm==0.0.25

conda activate snakemake
snakemake --version
8.20.6

Run Test Data Workflow

you are advised to run the following in a tmux or screen session

conda activate snakemake

# Set your cache dir for saving resources useful across other jobs, snakemake uses this when the `--cache` flag is set.

mkdir /fsx/resources/environments/containers/ubuntu/rnaseq_cache/
export SNAKEMAKE_OUTPUT_CACHE=/fsx/resources/environments/containers/ubuntu/rnaseq_cache/

# I set a partition relevant to my install, but if you specify nothing, you will get an error along the lines of <could not find appropriate nodes>.
snakemake --use-conda --use-singularity   --singularity-prefix /fsx/resources/environments/containers/ubuntu/ --singularity-args "  -B /tmp:/tmp -B /fsx:/fsx  -B /home/$USER:/home/$USER -B $PWD/:$PWD" --conda-prefix /fsx/resources/environments/containers/ubuntu/ --executor pcluster-slurm --default-resources slurm_partition=i64,i96,i192 --cache -p --verbose -k --max-threads 20000 --cores 20000 -j 14 -n   --conda-create-envs-only
- there seems to be a bug which requires you to run with  `--conda-create-envs-only` first...
- another bug with how snakemake detects max allowd threads per job limits the threads to the `nproc` of your head node.  Setting `--max-threads 20000 --cores 20000` gets around this crudely.

• Remove the -n flag, and run not in dryrun mode.
• -j sets the max jobs slurm will allow active at one time.
• Watch your running nodes/jobs using squeue (also, q cluster commands work, but not reliably and are not supported).
• note: it seems a bug in this example causes a few jobs to fail (investigating)

What Partitions Are Available?

Use sinfo to learn about your cluster (note, sinfo reports on all potential and active compute nodes. Read the docs to interpret which are active, which are not yet requested spot instances, etc). Below is what the daylily AWS parallel cluster looks like.

sinfo
PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
i8*          up   infinite     12  idle~ i8-dy-gb64-[1-12]
i64          up   infinite     16  idle~ i64-dy-gb256-[1-8],i64-dy-gb512-[1-8]
i96          up   infinite     16  idle~ i96-dy-gb384-[1-8],i96-dy-gb768-[1-8]
i128         up   infinite     28  idle~ i128-dy-gb256-[1-8],i128-dy-gb512-[1-10],i128-dy-gb1024-[1-10]
i192         up   infinite     30  idle~ i192-dy-gb384-[1-10],i192-dy-gb768-[1-10],i192-dy-gb1536-[1-10]
a192         up   infinite     30  idle~ a192-dy-gb384-[1-10],a192-dy-gb768-[1-10],a192-dy-gb1536-[1-10]

• As I look at this, it is possible that if unset, the partition will default to i8 in the output above. Maybe.

Budgets, and the --comment sbatch flag

I etensively make use of Cost allocation tags with AWS ParallelCluster in the daylily omics analysis framework $3 30x WGS analysis to track AWS cluster usage costs in realtime, and impose limits where appropriate (by user and project). This makes use of overriding the --comment flag to hold project/budget tags applied to ephemeral AWS resources, and thus enabling cost tracking/controls.

• To change the --comment flag in v0.0.8 of the pcluster-slurm plugin, set the comment flag value in the envvar SMK_SLURM_COMMENT=RandD (RandD is the default).