Description

Hello!

I would like the team to consider adding memory_per_cpu and memory_per_gpu variables/options to the workflow.toml file, as an override to the clusters default settings. It would be financially beneficial for some HPC users to be able to set the memory_per_cpu and memory_per_gpu , depending on the billing structure of some HPCs, which I will explain below.

I understand most HPCs auto-set the job memory per CPU-cores or GPUs. In this case, I clearly see why this is option/override is not needed.

There is a use case for GT's research HPCs where this is desired. On these HPC's, the research groups pay only on the CPU-core-hrs or GPU-hrs that are consumed (it is not free to use like a lot of other university HPCs), which is independent of the memory used. There is a default memory value per CPU-core or GPU if the users do not set a memory selection; however this is user selectable and not a fixed amount. This means that a user could technically use 1 CPU-core and all the memory on the machine and only be charged for 1 CPU-coretime_used. The same is true with GPUs, where you can use 1 GPU and all the memory on the machine and only be charged 1 GPUtime_used.

Maybe this charging method is not the standard, common, or idea. However, this is what it is for now, so this feature would be much appreciated. If this is not an option, I feel like some PIs may push back on using great tools like this, from solely a money saving objective. My goal is to promote great software like this as it will benefit the community by having efficient, repeatable, and reproducible workflows.

Given this HPC system's pricing structure is not likely to change anytime soon, it would be highly financially beneficial to have the memory_per_cpu and memory_per_gpu variables/options to the workflow.toml file. Research groups could could use this software and continue to minimize the research computing costs by not having to select many more CPU-cores above what they actually need just to get the memory. This would save ~5-10x the money for these types of jobs, if a lot of node memory is needed for a few CPU-cores.

There are several specific cases where researchers need like 1-4 cores and like 0.5-1TB of memory to load up things or 1 GPU and 0.5-1TB of memory.

Thank you for your consideration to add this feature.

Proposed solution

No response

Additional context

No response