niconiconi on Nostr: Just completed this cursed MPI + OpenMP setup. MPI processes are launched across CPU ...
Just completed this cursed MPI + OpenMP setup. MPI processes are launched across CPU sockets, OpenMP threads are launched within each socket. This is otherwise the standard, except that my MPI processes are also working directly on the same shared memory without any message passing... (MPI 3 already has threading support, just use that? No support for dynamic scheduling AFAIK. What about OpenMP's native nested parallelism? That was the plan, but I suspect GOMP's implementation has a performance problem that I need to verify, hence this Rube Goldberg parallel machine...) #hpc
Published at
2023-12-31 09:14:43Event JSON
{
"id": "d649d1207477e1c2c42aa718cdc4a1765b12498ec603135593e77a5b5808cfcf",
"pubkey": "26fdcfc5f6ff27a9926ef372554241c70b1042094ec7749aa301091e34d17aa0",
"created_at": 1704014083,
"kind": 1,
"tags": [
[
"t",
"hpc"
],
[
"proxy",
"https://mk.absturztau.be/notes/9nwuaurlaloauagf",
"activitypub"
]
],
"content": "Just completed this cursed MPI + OpenMP setup. MPI processes are launched across CPU sockets, OpenMP threads are launched within each socket. This is otherwise the standard, except that my MPI processes are also working directly on the same shared memory without any message passing... (MPI 3 already has threading support, just use that? No support for dynamic scheduling AFAIK. What about OpenMP's native nested parallelism? That was the plan, but I suspect GOMP's implementation has a performance problem that I need to verify, hence this Rube Goldberg parallel machine...) #hpc\n\nhttps://misskey-taube.s3.wasabisys.com/files/38a48161-7e5d-4a17-90d8-f1155efd67de.png",
"sig": "4c60f0e84062a6feb4485ddd234d8c402c4d4dbd2ec73ef7205bf1a62273df857cb0b41838d09212343ce10be3382082294b99731bcd8c816d040d81fd733d00"
}