Newsgroups: comp.parallel.mpi
From: Ben Edgington <ben@met.rdg.ac.uk>
Subject: Re: Q: derived data types and isend
Organization: University of Reading, U.K.
Date: 04 Oct 1996 09:44:02 +0100
Message-ID: <tg6pw2zruel.fsf@met.rdg.ac.uk>

s_straub@ira.uka.de (Matthias Straub) writes:
> I'm wondering how indexed derived data types effect
> performance, considering:
> You can't expect the communication-hardware to do the indexing,
> so MPI would probably copy the indexed data into a linear buffer and
> transfer this as usual. But what if immediate-sends are
> used ( which I plan to do)? There is no buffering - so does MPI have to 
> transfer the indexed data word by word?
> How about performance compared to 'manually' copying the indexed data 
> into a buffer via a do-loop?
> Any help appreciated.
> 
> 	Matthias Straub
> 
> BTW. hardware is a SP2 and a Power Challenge
> 

Don't know about SP2 and a Power Challenge, but...

For the Cray T3D Mike Ashworth from the DKRZ in Hamburg presented some work on
this at the 2nd European Cray MPP Workshop.  His slides and some discussion
can be found at http://www.epcc.ed.ac.uk/t3d/workshop/proceedings/Ashworth/

His conclusions are interesting: under some circumstances a local
memory-memory copy can be slower than passing a message between processors.
Therefore, more, shorter messages are more efficient than copying into
buffers; and similarly synchronous rather than asynchronous communications
(avoiding the buffer copy) may be more efficient.

Ben Edgington.
UGAMP, University of Reading

