Newsgroups: comp.parallel.mpi,comp.sys.super
From: dbader@Glue.umd.edu (David Bader)
Subject: Re: Buglist for IBM's SP-2 implementation of MPI
Organization: Institute for Advanced Computer Studies (UMIACS), Univ of Maryland, College Park
Date: 12 Aug 1996 11:07:08 -0400
Message-ID: <4unhas$s70@caddie.eng.umd.edu>

In article <4ukgr5$gnf@math2.cims.nyu.edu>,
Alan M. McKenney <mckenney@cims.nyu.edu> wrote:
>    Is there a bug list available for IBM's implementation
>of MPI for the SP-2?  (I *don't* mean MPICH.)  I am having
>a weird problem with sends never completing, and I would
>like to make sure I'm not running into some known problem
>in IBM's software before I spend a month looking for
>a bug in my code.

My group uses code on an IBM SP-2 built on top of IBM's MPI (was MPL
(was EUIH (was EUI/p ))). We were deadlocking in a very simple
situation which should be valid, when messages were over 4K * 4 bytes
long. (We've had conversations with other researchers who've run
across the same situation.) My actual code is more complex (based upon
Berkeley Active Messages), but essentially the culprit was a pair of
processors posting non-blocking receives followed by blocking sends to
each other. (No, I can't use a send-and-receive call because the
communication patterns are not known a priori.) My bug fix is specific
to active messages, but generally, the blocking sends are replaced
with non-blocking versions, and the code does other work while
checking the status of pending receives.

What was severely frustrating is that when I tried to recreate the
problem with IBM's parallel debugger, the situation did not re-occur.
I spent one week (an eternity for me to debug ;) developing a simple
IBM/MPI code which is not specifically invalid or dangerous according
to the man pages, but does in fact deadlock most of the time.

-david

David A. Bader, Ph.D. 
Institute for Advanced Computer Studies
A.V. Williams Building 
University of Maryland
College Park, MD 20742
Office: 301-405-6755   
FAX:    301-314-9658

Internet: dbader@umiacs.umd.edu
WWW:      http://www.umiacs.umd.edu/~dbader



