Hi! I'd like to talk about our recent experiments.

Achievement

We implemented the data copying scheme and several other optimization techniques in Polaris. Using the techniques, we automatically optimize the SPMD-style parallel programs which were translated from the original benchmark programs. The experiment results with optimized programs for 3 Perfect and 3 SPEC '95 benchmarks are shown in Graph 1 where we can see the significant improvement of the preliminary performance that we obtained before we optimize them.

New Techniques - to be implemented

We are currently developing and implementing more new techniques now, and apply the techniques by hand. The hand-optimized result is shown in Graph 2 below.

Future Plan

Eventually, we will implement these new techniques to further improve the speedups we obtained in Graph 1. In addition, we will conduct experiments on other Perfect Benchmarks and SPEC programs that are not shown here. As soon as we have the experiment results of those programs with the current optimization techniques, we will add the results to Graph 1, too. Also, we will extend our work on the T3D to other distributed memory machines such as the Convex Exemplar and networks of workstations.

Graph 1: Automatically Optimized Performance:

Graph 2: Automatically + Manaully Optimized Performance: