Newsgroups: comp.parallel
From: edemaine@ug.cs.dal.ca (Erik Demaine)
Subject: More messages -> slower?
Organization: Math, Stats & CS, Dalhousie University, Halifax, NS, Canada
Date: 5 Sep 1995 14:43:01 GMT
Message-ID: <42hnll$7ho@usenet.srv.cis.pitt.edu>

Hi Folks,

This is a very odd phenomena.  First:
	Parallel system:	IBM SP-2
	Number of nodes:	4 (only 2 used in tests)
	Message-passing:	IBM MPL or MPI (via MPICH)

The test:
	- calculate the megabyte/second rating of the switch
	- very simple calculations: round-trip for N messages, each of size
	  M.  Calculate total wall-clock time, and divide by N*M*2 (2 for the
	  round-trip part).  Synchronization between messages is made using
	  just a simple receive.

I found MPICH and MPL provided *very* similar results, so I'll only tell of
those for MPL.

I kept M constant (at 696 bytes).  When I jumped to around N=16000 from N=100,
I got *much* worse performance (around a factor of 10).  I've tried some
values in between, and it appears to be a linear increase.  (I haven't tried
a large amount of N's, but have gotten in-between speeds.)

Any suggestions as to why this is happening?

Thanks!
Erik
--
Erik Demaine        || edemaine@ug.cs.dal.ca  || edemaine@fx2800.dal.ca
edemaine@cs.dal.ca  || 01ERIK@ac.dal.ca       || edemaine@is.dal.ca
URL: http://ug.cs.dal.ca:3400/~edemaine/edemaine.html
--> A glob is a round substance of thick liquid. A blob is an unformed object.

