>>I hope that there is some performance gain that pays dividends for all
this
>>extra complication: we spend a lot of effort to keep track of Intel's
>>ever-changing API's for compilers, MKL and MPI.
At the moment, yes, the gain for me is significant (~20-30% for pmemd).
One of the things that impi seems to do is to set task affinity such that
the kernel scheduler is not constantly (and pointlessly) shuffling the MPI
threads around the different cores.
This is apparently present in oMPI 1.7.4:
http://www.theregister.co.uk/2014/02/07/openmpi_174_ships/
...but my cluster is still on 1.6.5 and the one that apt-get fetches for my
desktop is 1.4.3.
If its any help to anyone, I got tired of the configure2 logic around using
intel compilers a long time ago, and currently deal with the problem by
running a script like this before I ./configure -mpi intel:
export I_MPI_CXX=icc
export I_MPI_FC=ifort
export I_MPI_CC=icc
export I_MPI_F90=ifort
Josh
On 2 April 2014 03:26, David A Case <case.biomaps.rutgers.edu> wrote:
> On Tue, Apr 01, 2014, Dan Roe wrote:
>
> > My take on this is that the (now) long-standing behavior of Amber
> configure
> > has been to use mpicc/mpif90 as the designated compiler wrappers for MPI
> > compilation. The intel MPI 'mpiifort' etc. naming represents a deviation
> > from this, which is why I think it should be handled by a compiler flag,
> > much like the way '-crayxt5' deals with compiler wrappers on a cray
> system.
> > Also, just because a user has Intel MPI installed doesn't necessarily
> mean
> > they want to automatically use it for Amber. My vote is to make the Intel
> > MPI logic separate from the 'intel' compiler flag, and for it to require
> a
> > specific flag to configure.
>
> I agree with Dan here....looks like may need a new flag, say like
> -intelmpi.
> I hope that there is some performance gain that pays dividends for all this
> extra complication: we spend a lot of effort to keep track of Intel's
> ever-changing API's for compilers, MKL and MPI.
>
> ...dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Wed Apr 02 2014 - 01:30:03 PDT
