[Beowulf] Really efficient MPIs??

Christian Bell christian.bell at qlogic.com
Wed Nov 28 07:48:59 PST 2007


But the main point with MPI implementations, more than usual with
shared memory, is to run your application.

For 2 different MPI shared-memory implementations that show equal
performance on point-to-point microbenchmarks, you can measure very
different performance in applications (mostly at the bandwidth-bound
level).

Microbenchmarks assume senders and receivers are always synchronized
in time and report memory copy performance for memory copies that go
mostly through the cache.  Memory transfers that are mostly out of
cache are rarely tuned for or even measured.  

Microbenchmarks also never have the receivers actually consume the
data that's received or have senders re-reference the data sent for
computation.  The cost of these application-level memory accesses is
greatly determined by where in the memory hierarchy the MPI
implementation left the data to be computed on.  And finally, a given
implementation will have very different performance characteristics
on Opteron versus Intel, few-core versus many-core and point-to-point
versus collectives.

It's safe to assume that most if not all MPIs try to do something
about shared memory but I wouldn't be surprised if each of them can
top out on some performance curve on some specific system.  


    . . christian

On Wed, 28 Nov 2007, amjad ali wrote:

> Hello,
> 
> Because today the clusters with multicore nodes are quite common and the
> cores within a node share memory.
> 
> Which Implementations of MPI (no matter commercial or free), make automatic
> and efficient use of shared memory for message passing within a node. (means
> which MPI librarries auomatically communicate over shared memory instead of
> interconnect on the same node).
> 
> regards,
> Ali.

> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf


-- 
christian.bell at qlogic.com
(QLogic Host Solutions Group, formerly Pathscale)



More information about the Beowulf mailing list