Performance of Parallel Machines

Information Sciences Institute
University of Southern California

The figure illustrates our memory hierarchy Micro Benchmark, which characterizes the performance space of the memory hierarchy and the interconnection networks of different shared memory machines. Our group is developing new methods of analysing the performance of shared memory parallel machines, including an approach we call Micro Benchmarks. Micro Benchmarks are small benchmarks, each of which measures a particular aspect of the system under controlled conditions. We are also developing new methods of improving the performance of shared memory machines.

A recent advance is a new method of improving the scalability and performance of parallel applications on machines that permit prefetching. We have developed a method of adjusting the prefetch interval for cache lines based on the use of performance data, and ways of incorporating this method into compilers. Our initial tests shows that this method may improve performance by 30% for a range of 10 to 64 processors.

A description of the project, including the Micro Benchmark software distribution, the Micro Benchmark man page, results for specific machines, and relevant papers, may be found at http://www.isi.edu/div7/acs/acs.html