[mpich-discuss] MPI_Comm_accept() pins cpu core 100%

Dmitriy Lyubimov dlieu.7 at gmail.com
Mon May 9 11:08:48 CDT 2016


ok, thank you.

What implications does this have? does this reduce the solution
capabilities/performance in any specific way?

for now, i ended up just bumping process priority to the max priority for
the duration of accept call, which is still not great but i don't want to
mess with the build capabilities or ask people to install a custom build.

On Mon, May 9, 2016 at 6:35 AM, Balaji, Pavan <balaji at anl.gov> wrote:

>
> For now, the only workaround for this is to configure with
> --with-device=ch3:sock
>
>   -- Pavan
>
> > On May 6, 2016, at 5:49 PM, Dmitriy Lyubimov <dlieu.7 at gmail.com> wrote:
> >
> > Hi,
> >
> > I have a situation when a process my start up and execute
> MPI_Comm_accept(), while waiting for others to join, which may take some
> time.
> >
> > I have observed that all processes that wait in MPI_Comm_accept(), pin 1
> cpu core exactly 100%.
> >
> > Which is unfortunate in my situation since if i launch enough processes,
> i pin the entire machine node very quickly -- and it does do other things
> that run into a crawl.
> >
> > Is this a known problem? is there a workaround for this?
> >
> > I tried on ubuntu 14-04 LTS which installs mpich 3.0.4 by default but i
> also tested 3.2 Nov/2015 and they both exhibit this very behavior.
> >
> > Thank you very much in advance.
> > -Dmitriy
> > _______________________________________________
> > discuss mailing list     discuss at mpich.org
> > To manage subscription options or unsubscribe:
> > https://lists.mpich.org/mailman/listinfo/discuss
>
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20160509/9fe486d6/attachment.html>
-------------- next part --------------
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss


More information about the discuss mailing list