Newsgroups: comp.parallel.mpi
From: jonesy@cs.ucsb.edu (Chris R. Jones)
Subject: Low Latency MPI
Organization: Computer Science, UC Santa Barbara
Date: 4 Aug 1996 22:07:56 -0700
Message-ID: <4u3vjc$e58@zonker.cs.ucsb.edu>

               Low Latency MPI for Meiko CS-2 and ATM Clusters

I have just completed a thesis at the University of California, Santa Barbara
involving the implementation of MPI for the Meiko CS-2 and clusters of Fore
Systems ATM (Asynchronous Transfer Mode) connected SGI workstations.
An abstract of this thesis is included below:

ABSTRACT
--------

MPI (Message Passing Interface) is a standard providing a message passing
API motivated by portability and efficiency needs for
parallel and distributed computing.   In implementing MPI for any architecture,
it is desired that message passing is accomplished with low latency and high 
bandwidth.  

   This thesis presents an implementation of MPI for the Meiko CS/2,
a 64 node parallel computer, and 8 SGI workstations connected to an 
ATM (Asynchronous Transfer Mode) network and Ethernet.   MPI uses a tagged
message passing model, allowing the receiver to specify the type of message
it is ready to receive, requiring that the receiver and sender negotiate and
match sends and receives before transferring data to the receiver's buffer.  
This thesis presents a mechanism for safely sending data
to the receiver before sends and receives are matched, allowing for a 
decrease in latency while still preserving high bandwidth.  The overheads
involved in implementing this mechanism for the Meiko and ATM cluster are
examined.


WEB PAGE
--------
For more information on this implementation, and a copy of the thesis,
please see http://www.cs.ucsb.edu/~jonesy/mpi



