Evaluation of Forge: an interactive parallelization tool

摘要

The evaluation of Forge used five typical NAS applications. Analysis of the results revealed three critical areas in need of improvement. First, if tools parallelize loops without performing machine-specific optimizations, performance can be significantly degraded. For highly vectorized codes, the degradation is the most severe. Second, to help user find false dependencies, tool messages should be understandable by application scientists. Query facilities for variable usage and definition should be provided. To guide parallelization, tools must estimate granularity and overhead. Third, the benchmarks contain a large number of small to medium size loops which limit the maximum speedup obtainable by parallelization. Worse yet, they all contain serial loops whose bounds are proportional to problem size. When the size of these loops is large, the speedup can be independent of problem size. Merely converting sequential programs cannot satisfy NAS needs; one must support the design and implementation of parallel algorithms. Finally, issues in creating a future NAS parallel programming environment are discussed.