Wow - with that heavy an oversubscription, your performance experience certainly is reasonable. Not much you can do about it except reduce the oversubscription, either by increasing the number of computers or reducing the number of processes.
On Mar 1, 2012, at 1:33 AM, Pinero, Pedro_jose wrote: > Thank you for your fast response. > > 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 > > Regards > > Pedro > > > > >>On Feb 29, 2012, at 11:01 AM, Pinero, Pedro_jose wrote: > > >> I am using OMPI v.1.5.5 to communicate 200 Processes in a 2-Computers > >> cluster connected though Ethernet, obtaining a very poor performance. > > >Let me making sure I'm parsing this statement properly: are you launching > >200 MPI processes on 2 computers? If so, do >those computers each have 100 > >cores? > > >I ask because oversubscribing MPI processes (i.e., putting more than 1 > >process per core) will be disastrous to >performance. > > >> I have measured each operation time and I haver realised that the > >> MPI_Gather operation takes about 1 second in each >>synchronization (only > >> an integer is send in each case). Is this time range normal or I have a > >> synchronization >>problem? Is there any way to improve this performance? > > >I'm afraid I can't say more without more information about your hardware and > >software setup. Is this a dedicated HPC >cluster? Are you oversubscribing > >the cores? What kind of Ethernet switching gear do you have? ...etc. > > >-- > >Jeff Squyres > >jsquy...@cisco.com > >For corporate legal information go to: > >http://www.cisco.com/web/about/doing_business/legal/cri/ > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users
