Simplified expression of message-driven programs and quantification of their impact on performance

Communication latency and unpredictable delays in remote response times constitute significant impediments to achieving high performance on massively parallel computers. Message-driven execution is a promising technique to improve the performance of parallel computations by overlapping these delays with useful computation. This thesis explores message-driven execution for improving performance of parallel programs. Programming in message-driven style is difficult due to the split-phase transactions it requires and due to the nondeterministic arrival of messages. We developed language constructs to express dependences between messages and computations in order to simplify expression of message-driven programs. Predicting the performance of message-driven programs via simulations is difficult because the arrival order of messages changes as the machine characteristics change. We developed a trace-driven simulation methodology based on the those language constructs. We also conducted an extensive performance study of message-driven programs.U of I OnlyETDs are only available to UIUC Users without author permissio