Running on Alpine (CU Boulder / CURC)¶
pixi run run-alpine --configfile config/my-project.yml
Profile: cluster/alpine/config.yaml. CPU rules go to acpu with QOS
cpu-normal (24 h; use cpu-long for up to 7 d), GPU rules to aa100 with
gpu-normal (24 h; gpu-long 7 d). Account amc-general. QOS is set per
rule (qos: resource), so a rule can be moved to a long QOS by editing its
set-resources entry.
Site rules that matter here:
$HOMEis 2 GB. Clone the pipeline under/projects/$USER, and putoutput_directoryon/scratch/alpine/$USER(purged after ~90 days untouched: copy results worth keeping to/projectsor PetaLibrary).scripts/setup-env.shpoints the pixi, uv and singularity caches at scratch automatically when/scratch/alpineexists; exportPIXI_CACHE_DIR=/scratch/alpine/$USER/.cache/pixibefore the firstpixi installas well, since activation has not happened yet then.- Memory is coupled to CPUs on
acpu(3840 MB per CPU); the profile sizescpus_per_taskto covermem_mb. - Singularity needs
module load singularity. Setdnascent.singularity_setup: "module load singularity"in the project config, and runmodule load singularitybeforepixi run install-dnascent. - Builds (none are needed for this pipeline) belong on
acompile. - No dorado module:
pixi run setupinstalls it under the repo.
A whole PromethION flowcell of sup basecalling can exceed gpu-normal's 24 h
on one A100. Either split the run into samples, or move basecall /
basecall_run to qos: gpu-long. dorado.resume keeps the partial output
so a re-queued job continues where it stopped.