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

Wayne State University, Detroit MI

Investigators

Abstract

The increasingly common multi-core/many-core CPU architectures are effective for accelerating programsâ?? execution only when sufficient parallelism is maintained. For data-intensive programs the increased parallelism can severely compromise I/O efficiency: when a sequential program is parallelized, not only computations, but also the I/O operations associated with them, can be distributed among multiple processes. Because the execution order of the processes is usually determined by the scheduler at runtime, the relative progress of processes is nondeterministic and the order in which the processes issue their I/O requests is accordingly nondeterministic. It is this I/O nondeterminism that can substantially compromise I/O efficiency, and thus program performance, by negating the advantages of parallel execution. To address this problem the PI proposes a facility built either in the operating system kernel or in the runtime to streamline the service of I/O requests from different processes of a parallel program. The major distinction from conventional techniques for improving I/O performance is in the coordination of I/O request issuance, via I/O-aware process scheduling, to improve the locality of these requests for I/O-intensive multithreaded and MPI programs. If successful, the proposed research would introduce a disruptive technique for data-centric computing to effectively relieve the I/O bottleneck. This project also provides abundant research training opportunities for students, especially under-represented minority students in the southeast Michigan area, to help relieve the shortage of IT professionals with expertise in parallel computing and storage systems.