Jeff wrote: MPI_THREAD_MULTIPLE support barely works in v1.8. Why have it on by default, especially when there's a performance penalty?
I think the "barely works" state of threading support is a stronger argument for return to the 1.6.x behavior than PSM performance. Who knows what subtle bugs have become "enabled by default"? -Paul [Sent from my phone] On Nov 5, 2014 9:58 AM, "Jeff Squyres (jsquyres)" <jsquy...@cisco.com> wrote: > On Nov 5, 2014, at 12:03 PM, Joshua Ladd <jladd.m...@gmail.com> wrote: > > > I think this is a pretty significant change in behavior for a minor > release, Jeff. According to the interested parties: > > > > "I'm reporting a performance (message rate 16%, latency 3%) regression > when using PSM that occurred between OMPI v1.6.5 and v1.8.1. I would guess > it affects other networks too, but I haven't tested. The problem stems > from the --enable-smp-locks and --enable-opal-multi-threads options." > > Right, and we just fixed that problem. > > To be clear: the above report is from *before* we fixed the problem, not > from *after* we fixed it. After we fixed it, Intel confirmed that PSM > performance went back up to v1.6 levels. > > It was never in the plan to bring over the > enable-THREAD_MULTIPLE-beavhior-by-default functionality to the v1.8 > branch. It was a mistake that it got brought over. > > > So, this has been in effect the entire 1.8 series and now you want to > significantly alter the behavior. > > Put differently: I want to fix the performance bug. > > > I'm of the opinion that we should live with the mistake in the 1.8.x > series and provide users with clear guidance on the default behavior and > advice for disabling the threading support that is consistent across this > series. 1.9.x is a clean slate. > > I disagree. > > MPI_THREAD_MULTIPLE support barely works in v1.8. Why have it on by > default, especially when there's a performance penalty? > > -- > Jeff Squyres > jsquy...@cisco.com > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2014/11/16212.php >
