Tool-assisted performance measurement and tuning of UPC applications

The PGAS paradigm provides a shared-memory abstraction for programming distributed-memory machines. UPC, one of the more popular PGAS languages, improves ease of programming for the user, yet it also makes it difficult for performance analysis to cor-relate runtime behavior to program constructs. As efficient remote memory access is critical to performance, understanding the com-munication pattern can bring insight for performance diagnosis and tuning. In our study we develop a light-weight tracing mechanism to track remote accesses in UPC programs and correlate them to the program structures. Based on the profiling results, we also propose caching and access coalescing through automated code refactoring to improve communication efficiency for irregular codes. We an-alyze the communication performance of UPC applications on a cluster of SMPs, and show our tool-assisted optimizations achieve significant speedups over the original programs. 1