Add memory and cpu limits to processes launched through Apptainer / Singularity with a local executor

[OlivierCoen]

New feature

Hi,

As discussed in nf-core's Slack, it would be helpful to be able to limit the resources used by a process launched with Apptainer or Singularity when using the local executor.

Use case

Currently, there is a discrepancy between Docker and Apptainer / Singularity engine behaviours when running a pipeline locally (already explained in Nextflow documentation). Although docker containers are limited in terms of cpu and memory (using the task's cpu and memory), apptainer / singularity containers can take up all available cpus and memory available on the host machine.

This can also be a problem during CI testing with apptainer, since the executor is local.

Suggested implementation

In line with what was proposed by @bentsherman , a solution could be to add flags like:

executor.$apptainer.memoryLimits
executor.$apptainer.cpuLimits

to pass task.cpus and task.memory values down to the container parameters.

Thanks a lot for all the hard work!

[bentsherman]

I forgot that there is already a dedicated scope for apptainer settings:

apptainer.resourceLimits

[OlivierCoen]

Do you think that it could be possible to implement (in opt-in of course) the same behaviour for Apptainer / SIngularity as what is done for Docker when the executor is local?

I think that the main concern of people using apptainer locally / in a multi-user server (not HPC) is to have apptainer containers inherit from their parent task's cpus and memory

[nttg8100]

Hi @bentsherman, I think it was implemented before and reverted to not support it. The reason may be because the issue of the singularity feature consistency with "vm" flag between different os 0d04e0f#diff-3125bd0a7e17309b6868228888b237947dcb166b32fd999eb325ba0144c4d8ae. What do you think to simply add them back ?