[mpich-discuss] MPI scatter algorithms for BlueGene/P

[Jeff Hammond]

Those operations were never specially optimized on Blue Gene.  The
operations that were optimized heavily on Blue Gene were/are Barrier,
Bcast, Allreduce, and Alltoall(v).  Most of the other collective
optimizations were derivatives of that.  For example, one can
implement Allgather as a series of Bcasts.  And Reduce was optimized
as a side effect of Allreduce; in some cases, Allreduce was faster
than Reduce, which is counter-intuitive.

One optimization that Blue Gene had for Gather was to use Reduce with
BOR (on BG/P) and SUM (on BG/Q, for float types at least).  This
turned Gather of count=1 into Reduce of count=nproc (size of
communicator), but it was very faster for short messages.

There aren't many good optimizations for Scatter.  MPICH has most of
the generic ones, as you might expect.  It is possible that Scatter as
Alltoallv with only one non-zero in the count vector was faster than
MPICH, but this would surprise me.

In any case, all MPI collectives on BG benefit from very good
network-to-processor balance, minimal rendezvous (connectionless HW
and SW), trivial virtual-to-physical translation, good bisection
bandwidth of _electrically isolated_ torus networks, etc.

If you have a more specific question, I might have a better answer.

Best,

Jeff

On Sun, Mar 1, 2015 at 4:36 PM, Khalid Hasanov <xalid.h at gmail.com> wrote:
> Hello,
>
> First of all, I am not sure if this group is the right place for this
> question. If not I apologize for asking unrelated question.
>
> I read two papers about optimizing MPI collective communications on BG/L and
> BG/P.
> (Optimization of MPI Collective Communication on BlueGene/L Systems
> and MPI Collective Communications on The Blue Gene/P Supercomputer:
> Algorithms and Optimizations respectively). However, these two papers do not
> mention anything about MPI scatter and gather operations, I wonder if these
> two collective operations have been optimized for BlueGene or they use
> exactly the same algorithms from MPICH. Any reference appreciated.
>
>
> Best regards,
> Khalid
>
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss



-- 
Jeff Hammond
jeff.science at gmail.com
http://jeffhammond.github.io/
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss