Patrick Diehl

New paper: Evaluating HPX and Kokkos on RISC-V using an Astrophysics Application Octo-Tiger

Paper: https://dl.acm.org/doi/10.1145/3624062.3624230

Pre-print: https://arxiv.org/abs/2309.06530

#hpc #hpx #riscv

Evaluating HPX and Kokkos on RISC-V using an astrophysics application Octo-Tiger | Proceedings of the SC '23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis

ACM Other conferences
Nov 13, 2023 at 4:30amWeb
Show threadJanne MorenNov 13, 2023

@diehlpk
That was a fun one!

Though I wonder about the A64FX performance. In my own experience, for that CPU especially it really makes a huge difference exactly what compiler and math libraries you use.

GCC is pretty bad on it, while the ARM compiler (who received patches for A64FX from Cray/HPE) does much better, and about as well as the official Fujitsu compiler.

Show threadPatrick DiehlNov 13, 2023

@jannem

Thanks. We could try the armclang on Ookami. We tried the Fujitsu compiler using the native mode and clang mode on Supercomputer Fugaku. However, we experienced some compilation issues for our code.

We will try armclang after Supercomputing 23.

Show threadJanne MorenNov 13, 2023

@diehlpk
You're not actually testing the Fujitsu system. Me, I wouldn't bother. I do wonder how much it will matter to have a compiler well optimized for RISC-V though...

SC23 is fun so far. Looking forward to the exhibition tomorrow.

Show threadSriram RamkrishnaNov 13, 2023
@diehlpk are you using anything like #sycl here? The paper doesn't reference anything like that I know that kokkos let's you use different runtimes - was just curious. Pardon any ignorance on my part.
Show threadPatrick DiehlNov 16, 2023
@sri we are using the HPX backend in Kokkos. Alternatively one could use the #openmp backend of Kokkos. I am not sure if #SYCL supports #RISC-V.