Newsgroups: comp.parallel.mpi
From: srd320@cs.usask.ca (Satya Ranjan Dash)
Subject: Performance Issues in MPI
Keywords: Communication Overheads
Organization: University of Saskatchewan
Date: 29 May 1995 00:55:39 GMT
Message-ID: <3qb62b$at8@tribune.usask.ca>

Sorry folks I do not know why previous attempts of mine failed to
post the news item.

I  have been  fiddling around with  MPI-LAM.  I have  hit  a snag in
carrying  out  further  development. The issue  I  am facing is high
communication overhead.

On  my Ethernet LAN I have run a little test program to measure  the
performance for transferring double precision numbers (arrays) across
the MPI processor pool. I found that the cost for sending 1K of double
precision numbers is around 10-15 seconds on the  average and  in the
worst case about  25 seconds.  Clearly with such high communication 
overheads  serious computation on idling workstation clusters is difficult.

The intention of this news item is to explore how other users of MPI are
fairing in their work and whether either I am terribly goofing in my
measurements or there are some possible tecniques which I am not aware of.

The crucial point is that the underlying network appears to be OK. Since
times measured by the "tping" system call in LAM are effcient with an
order of 0.1-0.2 seconds for 4096 bytes while my apparently straight
forward program suffers from a very high cost.

To further test the LAN I broke  up large arrays into smaller size chunks.
This resulted in some improvement but nothing close to that measured
with "tping".  
 
I would very much appreciate ideas from other users and implementers.

--
Cheers.                                                         Dash

P.S.    If any one is interested in  either looking  in at my test
        program (to catch my leg) or to take measurements on their
        own;  pleae drop me  a line and I will mail back the code.
        It is written in an easily understandable way and produces
        neatly formatted outputs.  However, it is trivial.


