Compiler Support for Irregular Applications on DSM
For running irregular applications on distributed memory computers, current practice dictates the use of an inspector-executor model in conjunction with a programmer-developed message passing program. Chaos, from the University of Maryland, exemplifies this approach. A pure run-time software distributed shared memory (DSM) system may not provide as much performance, because of the lack of message aggregation.
I have collaborated with other researchers from Rice and the University of
Rochester to show that with a modest amount of compiler support for
aggregating messages, a software DSM system can extract just as much
performance as the inspector-executor approach. Our research is especially
interesting because we demonstrate that you can get high performance out of
a distributed memory machine while still using a shared memory programming
model.
Relevant Publications:
-
Compiler and Software Distributed Shared Memory Support for Irregular
Applications,
with Alan Cox, Sandhya Dwarkadas, Honghui Lu and Willy Zwaenepoel,
Proceedings of Principles and Practice of Parallel Programming (PPoPP), June 1997, Las Vegas, NV.
Back to the research summary page
Ramakrishnan Rajamony
E-mail: (MyLastName) at us.ibm.com [please do not e-mail me at cs.rice.edu]
Last updated at 18:33 CST on Thursday, January 22, 1998