[Beowulf] Performance characterising a HPC application

Richard Walsh rbw at ahpcrc.org
Mon Mar 26 15:01:45 PDT 2007

Greg Lindahl wrote:
> On Mon, Mar 26, 2007 at 03:58:59PM -0500, Richard Walsh wrote:
>> but of course aggregation is a legitimate optimization technique
>> because not all message patterns are of the Gups variety just as not
>> all memory references are absent locality.
> This is true, although I would call it more of an "ease of use" issue.
> Everyone in the MPI arena already knows they're supposed to send as
> big of a message as possible, so it's fairly rare to find
> high-performance MPI codes that see an improvement with message
    Agreed ... but ... without drifting to far off topic ...
> aggregation. In an ideal world the programmer wouldn't have to
> explicitly think about aggregation, it would just happen. But today's
> codes don't assume that.
    it becomes a requirement to obtain performance in more implicit
parallel programming
    models like UPC (unless you Balkanize it with programmed aggregation
via upc gets/puts or
    MPI-like UPC collectives).  Without a global memory vector
operation/instruction you
    cannot amortize latency using a pipeline as you can on the Cray
X1E.   Perhaps that is
    the ideal world that you are referring too ... ;-) ... although
aggregation has advantages over
    pipelining that stand on their.
> I only referred to GUPs as it's a widely available microbenchmark
> which is not gamed by this optimization. But message rate and
> streaming bandwidth are completely wrecked.
    Understood ...



Richard B. Walsh

Project Manager
Network Computing Services, Inc.
Army High Performance Computing Research Center (AHPCRC)
rbw at ahpcrc.org  |  612.337.3467

>  "The world is given to me only once, not one existing and one
>   perceived. The subject and object are but one."
>   Erwin Schroedinger

This message (including any attachments) may contain proprietary or
privileged information, the use and disclosure of which is legally
restricted.  If you have received this message in error please notify
the sender by reply message, do not otherwise distribute it, and delete
this message, with all of its contents, from your files.

More information about the Beowulf mailing list