Performance about eigenvalue solve

[Runfeng Jin]

Hi, I have seen the document "Implementing Singular Value and Symmetric/Hermitian Eigenvalue Solvers" and find there is only a single node compare for scalapck.  And the performace has a slight improve. What is the situation when scale to a large nodes like 128, 256 or 512? Does slate still obtain performance increase and how much is the performance?

We are currently wish to port our code to solve eigenvalue to GPU and wish to know whether now we can expect a large performance increase from slate? Or maybe we just stay to the scalapack now? 

Thank you!

Runfeng Jin

[Mark Gates]

Unfortunately, the eigensolver & SVD codes in SLATE do not yet scale to multiple nodes. The main bottleneck is the 2nd stage band-to-(bi or tri)-diagonal, which is currently multi-threaded but not parallelized with MPI.

Mark

Interim Director, Innovative Computing Laboratory (ICL)

Research Assistant Professor, University of Tennessee, Knoxville

https://icl.utk.edu/~mgates3/

[Runfeng Jin]

Hi, thank you for your answer! Do you have any suggested eigensolver for dense matrix and all eigenvalues that support multi-node and AMD GPU?  By the way, I found many solvers that support GPU but all single-node or sparse matrix. Is this still a pending task that diagonalize dense matrix for all eigenvalues in GPU and multi-node?

[Chris Hillenbrand]

ELPA has an experimental option to compile for AMD GPUs: https://github.com/marekandreas/elpa/blob/master/documentation/INSTALL.md

Does anyone know if the DPLASMA eigensolver works on multiple nodes with GPU? (It seems to be Nvidia only)

[Mark Gates]

The DPLASMA eigensolver is not GPU accelerated. I inquired with the DPLASMA developers, who said GPU acceleration covers potrf, trsm, trmm, gemm, LU-nopiv, and QR. GPU support is currently only CUDA.

Mark