cuking binary for HPCC version

[jjfarrell]

@vladsavelyev @lgruen

What are the steps to port cuking to an HPCC (os: AlmaLinux 8) with various nvida gpus? Is there an cuking binary or directions to compile to cuking c code? Or is the a docker image we can download with the binary that we could run locally?

We have about 150k samples were would like to try running cuking on our local HPCC system.

[lgruen]

Hi @jjfarrell! Thanks for reaching out and sorry for the delayed response. Neither @vladsavelyev nor I are working actively on this anymore.

cuKING should work fine in an HPC environment, but I'm not aware of anyone trying that before. There's shouldn't be too much required in terms of porting, as long as you can run Docker containers (e.g. usually using Singularity or similar, see Dockerfile) with GPU access.

The scripts assume that you have Hail set up, but that also shouldn't be too hard in an HPC environment.

As far as I recall, there are only two main components that currently assume Google Cloud as the environment, which would require changes:

• The main executable (cuking.cu) currently uses the Google Cloud Storage client to read inputs and write outputs. If you're using a mounted file system instead, that code should become a little simpler (search for gcs_client).
• The job scheduling (cloud_batch_submit.py) uses Cloud Batch for parallelization. You'd need to update that to use SLURM or similar.

Finally, note that cuKING assumes that you have plenty of GPU memory (NVIDIA A100 or later, ideally). If you have less memory, you'll probably not see as much of a speed-up.