You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
With improvements/changes in BLAS multithreading, Julia-level threads are competing with the BLAS threads and we don't see any improvement in most situations with use_threads=true. In fact, we sometimes see performance degradation. So let's drop multithreading for the bootstrap and simplify that code. We should instead develop an example for the docs showing how to optionally parallelize using Distributed for people with access to clusters / large computer.
NB: This doesn't prevent us from adding threading to various custom linear algebra methods, if those prove worthwhile. That said, I think that should be an internal performance optimization and not expose any changes in the external API.
The text was updated successfully, but these errors were encountered:
With improvements/changes in BLAS multithreading, Julia-level threads are competing with the BLAS threads and we don't see any improvement in most situations with
use_threads=true
. In fact, we sometimes see performance degradation. So let's drop multithreading for the bootstrap and simplify that code. We should instead develop an example for the docs showing how to optionally parallelize using Distributed for people with access to clusters / large computer.NB: This doesn't prevent us from adding threading to various custom linear algebra methods, if those prove worthwhile. That said, I think that should be an internal performance optimization and not expose any changes in the external API.
The text was updated successfully, but these errors were encountered: