search

N8-CIR-Bede / documentation

Documentation for the N8CIR Bede Tier 2 HPC faciltiy
https://bede-documentation.readthedocs.io/en/latest/
7 stars 11 forks source link

BLAS/LAPACK clarification #65

Open loveshack opened 2 years ago

loveshack commented 2 years ago

Note that ESSL is an incomplete implementation of LAPACK, so you may need to back it with the reference one or openblas. While we're not supposed to be concerned with CPU compute, and I'd rather advocate free software, ESSL is significantly faster than openblas on level 3 BLAS. More importantly, it has CUDA support; that doesn't perform as well as I expected, but was better than nvblas in my tests.

Fehings commented 2 years ago

Would it be sufficient to link to the documentation that openBLAS and ESSL provide, and add a comment about the lack of CUDA support on the side of openBLAS, or is there further clarification needed?

loveshack commented 2 years ago

You wrote:

Would it be sufficient to link to the documentation that openBLAS and ESSL provide, and add a comment about the lack of CUDA support on the side of openBLAS, or is there further clarification needed?

Apologies for the delay. The main point was that ESSL shouldn't be advertised as LAPACK without qualification. (I wonder why they don't just add vanilla versions of the missing routines.)

I don't remember what doc there is on ESSL -- not enough to help me get good GPU GEMM performance anyway. I don't think there is much of consequence for OpenBLAS; it's just BLAS, and the only control you have is threading and forcing the micro-architecture through the environment. It might be worth recommending what to use for performance, at least: essl v. openblas, essl v. nvblas.

There doesn't seem much interest in multi-node use, but if anyone needs distributed linear algebra a la ScaLAPACK, SLATE and COSMA would be worth investigating for GPU support. I might give that a go sometime.

loveshack commented 2 years ago

Somewhere I have plots of essl and openblas performance on level 3 BLAS if it's of any interest.