|
Custom Collective Communication
Scott B. Baden
University of California, San Diego
Hosted by Nicholas Karonis
10:30 AM, June 18, 1999
Building 221, Room A-261
|
| Abstract |
Message passing is a popular approach for programming multicomputer
architectures, and application programmer interfaces such as MPI have enabled significant
scientific discovery using computer simulation techniques. However, message passing has
been criticized as too low level; it hides neither policy decisions nor application- and
platform-dependent optimizations. In this talk, Ill describe a framework for expressing
communication algorithms that separates correctness concerns from policy decisions and
optimizations that affect performance. This framework is part of the KeLP system developed
at UCSD. Using a set of geometric meta-data types along with C++ encapsulation mechanisms,
I'll show how to succinctly express a variety of customized communication algorithms
arising in the important class of block structured applications. I'll present performance
data demonstrating the benefits of such abstraction, which are able to overcome common
defects or omissions in MPI and its implementations on SMP-based multicomputers. |
