FS#68031 - [openmpi] enable CUDA support
Attached to Project:
Arch Linux
Opened by Jakub Klinkovský (lahwaacz) - Monday, 28 September 2020, 14:46 GMT
Last edited by Levente Polyak (anthraxx) - Sunday, 21 March 2021, 12:07 GMT
Opened by Jakub Klinkovský (lahwaacz) - Monday, 28 September 2020, 14:46 GMT
Last edited by Levente Polyak (anthraxx) - Sunday, 21 March 2021, 12:07 GMT
|
Details
Using GPU codes on HPC clusters or workstations often
requires using so-called "CUDA-aware" MPI in order to avoid
extra copies when working with data allocated on the GPU. In
OpenMPI, the CUDA support can be enabled with
"--with-cuda=/opt/cuda". See
https://www-lb.open-mpi.org/faq/?category=buildcuda
for details.
It would be nice to enable CUDA support in the Arch package. As far as I can tell, "cuda" can be added just to makedepends and optdepends, since OpenMPI has a dynamic plugin architecture and building the CUDA plugins should not affect the behaviour on systems where CUDA is not installed. Adding the --with-cuda flag leads to the compilation of these extra files: usr/lib/openmpi/libmca_common_cuda.so usr/lib/openmpi/libmca_common_cuda.so.40 usr/lib/openmpi/libmca_common_cuda.so.40.20.0 usr/lib/openmpi/openmpi/mca_btl_smcuda.so usr/lib/openmpi/openmpi/mca_coll_cuda.so usr/lib/openmpi/openmpi/mca_rcache_gpusm.so usr/lib/openmpi/openmpi/mca_rcache_rgpusm.so usr/share/openmpi/help-mpi-btl-smcuda.txt usr/share/openmpi/help-mpi-coll-cuda.txt usr/share/openmpi/help-mpi-common-cuda.txt |
This task depends upon
Closed by Levente Polyak (anthraxx)
Sunday, 21 March 2021, 12:07 GMT
Reason for closing: Implemented
Additional comments about closing: 4.0.5-3
Sunday, 21 March 2021, 12:07 GMT
Reason for closing: Implemented
Additional comments about closing: 4.0.5-3
- Make CUDA-aware support dynamically load libcuda.so so CUDA-aware
MPI library can run on systems without CUDA software.
https://raw.githubusercontent.com/open-mpi/ompi/v4.0.x/NEWS