Post #2558211
2026-04-30 22:06 UTC
@trs@metasocial.com @glyph@mastodon.social That too, but I'm seeing this with an otherwise single-threaded program. Turning off the OpenBLAS thread pool speeds things up.
Replies (2)
-
@dalke@toots.nu 2026-05-01 04:09
@itamarst@hachyderm.io @trs@metasocial.com @glyph@mastodon.social When I used OpenMP to parallelize my code I found that single-threaded processing became slower. I ended up using some C macros shenanigans to double-include the relevant C code, once for the single-thread case (no OpenMP) and once for multiple threads (OpenMP). If you do OMP_NUM_THREADS=1 then it avoids using OpenMP completely.
-
@kroekermartin.bsky.social@bsky.brid.gy 2026-05-01 10:18
How many cores do your systems usually have ? OpenBLAS's all-or-nothing approach to multithreading is likely to hurt for smallish problems on big machines, though there are some mitigations in place now specifically for ARM servers.