CSR: Small: Adaptively Applying Data-Driven Execution Mode to Remove I/O Bottleneck for Data-Intensive Computing

点击翻译按钮获取中文摘要

The increasingly common multi-core/many-core CPU architecturesare effective for accelerating programsâ?? execution only whensufficient parallelism is maintained. For data-intensiveprograms the increased parallelism can severely compromise I/Oefficiency: when a sequential program is parallelized, not onlycomputations, but also the I/O operations associated with them, canbe distributed among multiple processes. Because the executionorder of the processes is usually determined by the scheduler atruntime, the relative progress of processes is nondeterministicand the order in which the processes issue their I/O requests isaccordingly nondeterministic. It is this I/O nondeterminism thatcan substantially compromise I/O efficiency, and thus programperformance, by negating the advantages of parallel execution.To address this problem the PI proposes a facility built eitherin the operating system kernel or in the runtime to streamlinethe service of I/O requests from different processes of aparallel program. The major distinction from conventionaltechniques for improving I/O performance is in the coordinationof I/O request issuance, via I/O-aware process scheduling, toimprove the locality of these requests for I/O-intensivemultithreaded and MPI programs.If successful, the proposed research would introduce a disruptivetechnique for data-centric computing to effectively relieve theI/O bottleneck. This project also provides abundant researchtraining opportunities for students, especially under-representedminority students in the southeast Michigan area, to help relievethe shortage of IT professionals with expertise in parallelcomputing and storage systems.