FIFTH
ANNUAL
ASAP '97 WORKSHOP
|
FIFTH
ANNUAL |
Automated Application |
William J. Kostis, Gary E. Adams, Robert C. Durie, and Adam W. Bojanczyk Cornell University School of Electrical Engineering Ithaca, NY 14853 Abstract Development of high-perfomance radar and sonar signal processing applications continues to be a daunting task. Performance optimization, code reusability, and algorithm verification are each goals requiring a substantial effort in software engineering. In this paper, we describe our Parallel Program Factory (PPF), a development system designed to address these issues. An illustrative example application is taken from Space-Time Adaptive Processing (STAP). Achieving optimal application performance for a variety of problem sizes requires careful choices of both algorithms and, indeed, particular implementations of these computational "modules." The communication demands imposed by the data distribution requirements of each implementation further complicates the task. For example, it may be beneficial to use a data distribution that is inefficient for one module in order to increase the performance of another, more compute-intensive task. These relationships are of particular interest as the performance of many complex signal processing applications can be communication-limited. While applications may be portable across distributed memory platforms, performance frequently is not. Our approach relies upon libraries of performance benchmarks and models for each computational module implementation on each platform. Similar data is maintained for all data distribution modules. The top level application description can be specified using the Khoros dataflow environment. An application designer need only select the sequence of computational modules for the PPF to assemble a near-optimal software solution for the designated problem size and platform. In this way, machine performance characteristics, complex data distributions, and the effects of memory hierarchy are all shielded from the developer. In addition, a near optimal solution for a different problem size or platform can be generated automatically. |
Direct comments and questions to: webmaster@ll.mit.edu
© MIT Lincoln Laboratory. All rights reserved.
