Performance of the Efficient Data-Driven Evaluation Scheme

The Efficient Data-Driven Evaluation Scheme is a two-phase strategy for optimizing the execution time of programs in the IF1 dataflow format. The first phase involves adding extra nodes and edges to the IF1 program graph at compile-time. In the second phase, the new nodes and edges facilitate the determination of "useless" computations (i.e., computations which do not contribute to the final results) at run-time and prevent these computations from being executed. By focusing on the execution of potentially "useful" computations, architectural models with limited processors can be used more effectively, and execution times may be reduced. In this paper, we briefly review the Efficient Data-Driven Evaluation Scheme and describe performance results of the scheme on a large set of benchmarks which include the Livermore Loops. Our experiments compare the Efficient Data-Driven Evaluation Scheme to eager and lazy compilation for these benchmarks and characterize programs for which our scheme may be used effectively as a compiler optimization.