User Forum of Software BASEMENT

ClaireGilder

User

From: Missoula, Montana, US

Registered: 2023-11-19

Posts: 7

Parallelizing on multiple GPU

I have access to an 8 GPU cluster for running my model. Is there a flag I can set in my .yaml file that I use to launch the job to have the model take advantage of more than one GPU? Right now, the only flag I have is --backend cudaC in the BMv4_simulation command.

Additionally, even when I launch the model on a single GPU, it is only hitting 70% utilization and using roughly 9GB of vRAM, but i have 96GB available. Any suggestions on how to use the hardware I have more efficiently and hopefully speed up my model a bit? Are there any settings I can set in the simulation.json?

Thanks for any help!

Matthias Bürgler

Developer

From: Zürich

Registered: 2019-04-04

Posts: 188

Re: Parallelizing on multiple GPU

Hi Claire,

At the moment, BASEMENT does not support running a single simulation on multiple GPUs, so there is no additional YAML or simulation.json setting for this.

GPU memory usage is determined by the problem size and enabled features, such as the number of mesh cells and sediment grain classes. Having more GPU memory available than the simulation requires will therefore not improve performance. Similarly, GPU utilization may remain below 100% depending on the model setup and computational workload.

Best regards,
Matthias