Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: Re: [OMPI users] Very slow MPI_GATHER
From: Jeffrey Squyres (jsquyres_at_[hidden])
Date: 2012-03-01 06:12:16


On Mar 1, 2012, at 3:33 AM, Pinero, Pedro_jose wrote:

> I am launching 200 light processes in two computers with 8 cores each one (Intel i7 processor). They are dedicated and are interconnected through a point-to-point Gigabit Ethernet link.
>
> I read about oversubscribing nodes in the open-mpi documentation, and for that reason I am using the option
>
> - Mca mpi_yield_when_idle 1

That's still going to give you terrible performance.

Open MPI was designed to run basically at one process per processor (usually a core). The easiest reason to cite here is that Open MPI busy-polls while blocking for message passing progress. The yield_when_idle option *helps* (in some versions of Linux, at least), but it doesn't change that fact that MPI processes will be extremely aggressive in clamoring for CPU cycles.

-- 
Jeff Squyres
jsquyres_at_[hidden]
For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/