Performance Evaluation and Optimization for Scientific Kernels

Sponsor: Silicon Graphics Inc.

This project investigates a range of algorithmic, architectural, and system software issues related to performance evaluation and optimization on SGI machines. Our ongoing research has focused on the use of threaded APIs for optimized parallel kernels in linear algebra on the Origin 2000. We have achieved parallel efficiencies in excess of 85% on reduced-error multipole techniques for up to 32 processors using POSIX threads. Similar performance has also been demonstrated for multilevel solvers used in modeling incompressible particulate fluid flows. The proposed project extends this research in several directions:

• Study the impact of heterogeneous communication layers (Origin 2000 multiclusters of over 64 processors) on performance of threaded APIs. Specifically, we are investigating the impact of higher latency meta-routers and the ability of threads to mask additional latency.
• Using performance data (cache misses, synchronization penalty, context switches) to optimize scheduling for both uniprocessor and parallel performance. This study focuses on exploiting the iterative nature of most applications to enhance data access locality.
• Investigating architectural implications of parallel overheads. Specifically, identify machine parameters that most critically impact the performance of various kernels and techniques for alleviating performance bottlenecks.
• Algorithmic improvements for minimizing overheads. Often, a more expensive algorithm is capable of yielding better processor performance. Investigating such algorithms may be key to faster problem solving.

The issues addressed in this proposal impact critical design decisions for hardware, software, and application engineers. In addition to methodologies for improving performance, the project will result in optimized kernels for target applications in sparse linear algebra and n-body methods.