A Systematic Measurement of the Influence of Non-Uniform Cache Sharing on the Performance of Modern Multithreaded Programs

Most modern Chip Multiprocessors (CMP) feature shared cache on chip, whose influence on the performance of multithreaded programs, unfortunately, remains unclear due to the limited coverage of the deciding factors in prior studies. In this work, we conduct a systematic measurement of the influence using a recently released CMP benchmark suite, PARSEC, with a spectrum of factors considered. The measurement shows that, contrary to previous observations in multiprogramming environments and server programs, shared cache has insignificant influence on most of the programs, regardless of variations in input, thread assignment, and other factors. It suggests the mismatch between current multithreaded applications and CMP architectures, the limited potential of thread co-scheduling on those programs, and the necessity of program-level transformations for effective exploitations of modern CMP architectures. We apply a shared-cache-aware transformation to a data-mining program and produce over 30% speedup, preliminarily verifying the potential of program-level transformations.