[gmx-developers] rvec structure

[Nathan Moore]

Dear Berk, David,

Well, the implementation with collectives is marginally better.  It looks
from the trace files though that there's still considerable time spent in
MPI_Wait (often more time in Wait than in the collective).

Did an alternate version of gmx_sumf using MPI collectives (function
defined in network.c) ever get written?  Any insight on where else the
MPI_wait might be occurring?  So far I've only looked at the dpcc and
villin benchmarks.

regards,

Nathan Moore

 > Hi,
>
> You don't need a temp array.
>
> My test code is in CVS revision 1.8 of mvxvf.c:
>
>  >   static int *recvcounts=NULL,*displs;
>  >   int i;
>  >
>  >   if (recvcounts == NULL) {
>  >     snew(recvcounts,nsb->nprocs);
>  >     snew(displs,nsb->nprocs);
>  >     for(i=0; i<nsb->nprocs; i++) {
>  >       recvcounts[i] = nsb->homenr[i]*sizeof(x[0]);
>  >       displs[i]     = nsb->index[i]*sizeof(x[0]);
>  >     }
>  >   }
>  >   MPI_Allgatherv(arrayp(x[nsb->index[nsb->pid]],nsb->homenr[nsb->pid]),
>  >                MPI_BYTE,x,recvcounts,displs,MPI_BYTE,MPI_COMM_WORLD);
>
> It also contains the force summing, just one like is enough for this:
> MPI_Allreduce(f,f,nsb->natoms*DIM,mpi_type,MPI_SUM,MPI_COMM_WORLD);
>
> Berk.
>
> _______________________________________________
> gmx-developers mailing list
> gmx-developers at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-developers
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-developers-request at gromacs.org.
>