[Beowulf] Re: Naive question: mpi-parallel program in multicore CPUs

Herbert Fruchtl herbert.fruchtl at st-andrews.ac.uk
Wed Oct 3 01:28:05 PDT 2007

What is really difficult with MPI is data distribution. A lot of 
applications are parallelised using replicated data. That's fine if you 
are CPU bound, but it you are limited by the amount of memory per 
processor, a shared memory approach (which is the default with OpenMP) 
is the easiest way of using all the memory.

MPI may also add a lot of overhead if you parallelise inner loops, which 
is easy and cheap with OpenMP. OTOH, coarse-grain parallelism with 
OpenMP is difficult; MPI is usually more suitable here. It depends on 
your application, and you may find candidates for both approaches in the 
same application.


beowulf-request at beowulf.org wrote:
> Send Beowulf mailing list submissions to
> 	beowulf at beowulf.org
> To subscribe or unsubscribe via the World Wide Web, visit
> 	http://www.beowulf.org/mailman/listinfo/beowulf
> or, via email, send a message with subject or body 'help' to
> 	beowulf-request at beowulf.org
> You can reach the person managing the list at
> 	beowulf-owner at beowulf.org
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of Beowulf digest..."
> Today's Topics:
>    1. Re: Naive question: mpi-parallel program in multicore CPUs
>       (Larry Stewart)
> ----------------------------------------------------------------------
> Message: 1
> Date: Tue, 02 Oct 2007 14:52:21 -0400
> From: Larry Stewart <larry.stewart at sicortex.com>
> Subject: Re: [Beowulf] Naive question: mpi-parallel program in
> 	multicore CPUs
> To: Li at mx2.buaa.edu.cn, Bo <libo at buaa.edu.cn>
> Cc: beowulf at beowulf.org, Kwan Wing Keung <hcxckwk at hkucc.hku.hk>
> Message-ID: <47029365.8090606 at sicortex.com>
> Content-Type: text/plain; charset="gb2312"
> The question of OpenMP vs MPI has been around for a long time,
> for example:
> http://www.beowulf.org/archive/2001-March/002718.html
> My general impression is that it is a waste of time to convert from pure
> MPI to
> a hybrid approach. For example:
> www.sc2000.org/techpapr/papers/pap.pap214.pdf
> On the other hand, here's a fellow who got a 4X speedup by going to hybrid:
> www.nersc.gov/nusers/services/training/classes/NUG/Jun04/NUG2004_yhe_hybrid.ppt
> My own view is that with a modern cluster with fast processors and with
> inter-node communications not
> that much slower than a cache miss to main memory, the unified MPI model
> makes more sense, but
> there are many many papers arguing about this topic.
> -L
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL: http://www.scyld.com/pipermail/beowulf/attachments/20071002/567c759a/attachment-0001.html
> ------------------------------
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf
> End of Beowulf Digest, Vol 44, Issue 4
> **************************************

Herbert Fruchtl
EaStCHEM Fellow
School of Chemistry
University of St Andrews

More information about the Beowulf mailing list