[gmx-users] Launch hybrid MPI/openMP run on multiple nodes

Mark Abraham mark.j.abraham at gmail.com
Tue Jan 19 13:49:17 CET 2016


Hi,

No, that is a per-rank limit.

Mark

On Tue, Jan 19, 2016 at 11:38 AM Chunlei ZHANG <chunleizhang.pku at gmail.com>
wrote:

> Dear Sizilard and Mark,
>
> Thank you very much for the explanations.
> I found a line in the log file as:
>
> OpenMP support:     enabled (GMX_OPENMP_MAX_THREADS = 32)
>
> I wonder whether this setup limits the number of OpenMP threads and hence I
> can not launch more than 32 threads on multiple nodes.
>
> Best,
> Chunlei
>
> 2016-01-15 22:01 GMT+08:00 Szilárd Páll <pall.szilard at gmail.com>:
>
> > Hi,
> >
> > Your job scheduler and/or MPI launcher is most likely to blame. The fact
> > that mdrun warns about the logical core (hardware thread) mismatch means
> > that the OpenMP runtime thinks that you should be using one thread per
> > rank. This typically means that the MPI launcher or job scheduler set an
> > affinity mask for each rank and likely mdrun also skipped pinning threads
> > because of this.
> >
> > This should however alone not cause nodes to be empty, but rather it
> would
> > cause ranks to overlap and run on the same core.
> >
> > In any case, you should pass the correct ranks/node threads/rank settigns
> > to your launchers; to ensure correct rank placement you'll have to either
> > set up affinities through the scheduler or let mdrun do it.
> >
> > Cheers,
> > --
> > Szilárd
> >
> > On Fri, Jan 15, 2016 at 9:25 AM, Chunlei ZHANG <
> chunleizhang.pku at gmail.com
> > >
> > wrote:
> >
> > > Dear GMX developers and users,
> > >
> > > I have a cluster of 24 nodes, each having two 10-core intel CPUs.
> > > Gromacs 5.1 is compiled by using intel mpi (version 5.1.1) and mkl.
> > >
> > > I can successfully run a simulation by using pure MPI (480 MPI
> > processes).
> > > But the performance is not good and the log file of mdrun suggests
> using
> > > fewer MPI processes.
> > > I try to launch 240 MPI processes, each using 2 openMP threads, by the
> > > command:
> > > mpirun -ppn 20 -np 240 gmx_mpi mdrun -ntomp 2
> > >
> > > But only a fraction of the nodes are running mdrun and the log file
> says:
> > >
> > > Number of logical cores detected (20) does not match the number
> reported
> > by
> > > OpenMP (1).
> > > Consider setting the launch configuration manually!
> > >
> > > Does anyone know how to solve this problem?
> > > Thanks in advance.
> > >
> > > Best,
> > > Chunlei
> > > --
> > > Gromacs Users mailing list
> > >
> > > * Please search the archive at
> > > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > posting!
> > >
> > > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > >
> > > * For (un)subscribe requests visit
> > > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > > send a mail to gmx-users-request at gromacs.org.
> > >
> > --
> > Gromacs Users mailing list
> >
> > * Please search the archive at
> > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > posting!
> >
> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >
> > * For (un)subscribe requests visit
> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > send a mail to gmx-users-request at gromacs.org.
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.


More information about the gromacs.org_gmx-users mailing list