SPARTAN: A software tool for Parallelization Bottleneck Analysis

The multicore era has brought to an end the trend of doubling single-thread performance with each generation of microprocessors. To continue scaling programmers must parallelize existing single-threaded applications. A lack of tools makes identifying program features and data-dependence relationships that bottleneck parallel performance a cumbersome ad-hoc task. This paper presents SPARTAN, a tool that analyzes single-threaded applications, and points out the data-dependences that are likely to pose bottlenecks to parallel performance. In addition SPARTAN ranks the bottlenecks by giving an estimate of their impact on performance, so that these can be fixed in an appropriate order. We validate SPARTAN by showing improved parallel speedup on a benchmark application when the identified bottlenecks are removed.