SLATE usability on multi-GPU and multi-node

[Federico Cipolletta]

Good morning,

I am evaluating the usage of the SLATE library to solve the generalized eigenvalue problem on dense matrices in the CARIDDI code, which is a state-of-the-art tool to compute the electromagnetic contributions of passive structures around a given geometry. The current implementation of the code relies on the full spectrum of the eigenvalues of given matrices, with size directly proportional to the DoF of the geometrical model. The code currently implements ScaLAPACK, and we are testing valid alternatives that would allow the exploitation of modern HPC architectures, with a particular focus on hybrid CPU-GPU ones.

I would like to know if the eigensolver provided by SLATE can be efficiently used on multiple GPUS and multiple nodes. I noticed this discussion on the web

https://groups.google.com/a/icl.utk.edu/g/slate-user/c/iBLG-wdaG4E

where that possibility seems to be excluded. Given that the latest message is from March 2024, I would like to know if there have been improvements in that direction and if there is any reference reporting those improvements.

Best Regards,

Federico Cipolletta, PhD
Senior Scientific Researcher
Barcelona Supercomputing Center

[Mark Gates]

Hi Federico,

Unfortunately, there haven't been any updates to SLATE's Eig or SVD codes, due to a lack of funding.

The Eig and SVD can use multiple GPUs within a single MPI process. The 2nd stage doesn't scale with multiple MPI processes; it should scale with CPU threads within one MPI process.

Mark

[Federico Cipolletta]

Hello Mark,

Thank you for the precise response and for the update on the status of the work.

Best Regards,

Federico Cipolletta