Published June 4, 2024
| Version GPU_v1
Software
Open
fmihpc/vlasiator: Vlasiator GPU pre-release v1
Creators
- 1. University of Helsinki
- 2. @UCL-ARC
- 3. CSC - IT Center for Science
Description
This pre-release is a functional first release of Vlasiator with support for NVIDIA GPUs via CUDA and AMD GPUs via HIP/ROCm. The same branch also supports building for CPUs just as before. Slight differences in output show up as a result of differing summation orders in e.g. moments reduction, but testpackage tests all pass with acceptable numerical diffs.
Known issues:
- Block adjustment in acceleration can still be slow.
- VDF-based data reducers are somewhat slow.
- AMD performance is low and requires a mempool for operations (e.g. https://github.com/sfantao/vlasiator-mempool). Planned improvements include support for more robust memory management libraries.
- Field solver and ionospheric grid solver have not yet been ported to GPUs.
- GPU RDMA may not work on all systems, leading to poor MPI performance.
- CUDA
compute-sanitizer
reports several invalid context errors on launch; these can be ignored.
This introduces the Hashinator library (https://github.com/fmihpc/hashinator) as a required submodule. Make sure to initialize and pull it as well. GPU compilation should use a GPU-compatible compiler for all objects; see e.g. MAKE/Makefile.mahti_cuda
for instructions on how to set up a compiler.
Files
fmihpc/vlasiator-GPU_v1.zip
Files
(13.9 MB)
Name | Size | Download all |
---|---|---|
md5:8e2ac5239b916e863eedc3fe809c0d43
|
13.9 MB | Preview Download |
Additional details
Related works
- Is supplement to
- Software: https://github.com/fmihpc/vlasiator/tree/GPU_v1 (URL)