From: sriram@cs.utexas.edu (Sriram S. Rao)
Newsgroups: comp.parallel.mpi
Subject: Help---Need MPI applications
Date: 21 Mar 1999 21:47:36 -0600
Organization: CS Dept, University of Texas at Austin
Message-Id: <7d4ego$rfb$1@grampa.cs.utexas.edu>
Xref: ukc comp.parallel.mpi:4791


Hi:
	I have developed a middleware for providing
fault-tolerance and I have integrated it with MPICH.
The middleware uses message logging protocols
in conjunction checkpointing for providing fault-tolerance.
I have used NPB 2.3 benchmark suite for
some preliminary evaluation.  The problem with
those applications is that communication is
synchronous---in the sense that all processes compute
for nearly the same amount of time in a loop and
then exchange messages.

To evaluate the performance of logging and checkpointing protocols 
I am interested in obtaining other types MPI applications from any of you
to do a performance evaluation.   Specifically, I am 
interested in applications in which communication
is asynchornous and computational load on each
process is different.  Such applications will demonstrate the
benefits of logging messages.  Consider the following
scenario in which there are at least two process failures
in a distributed application.  Say that, process p, q fail.
  --> if all the messages that q sent to p before failing
      were logged on disk by p, then during p's recovery,
      p can replay those messages *without* incurring
      any blocking.  (During normal execution, p would have
      had to block waiting for q to send the message)
  --> if no messages were logged, then recovering p will have
      to wait for q to recover and send it messages---this is
      similar to failure-free execution.

If any of you can point to me to some applications, it
would be really great.

My email address: sriram@cs.utexas.edu

Thanks.

Sriram Rao
Department of Computer Sciences,
UT-Austin