Automatic data aggregation for software distributed shared memory systems

[摘要] Software Distributed Shared Memory (DSM) provides a shared-memory abstraction on distributed memory hardware, making a parallel programmer;;s task easier. Unfortunately, software DSM is less efficient than the direct use of the underlying message-passing hardware. The chief reason for this is that hand-coded and compiler-generated message-passing programs typically achieve better data aggregation in their messages than programs using software DSM. Software DSM has poorer data aggregation because the system lacks the knowledge of the application;;s behavior that a programmer or compiler analysis can provide. We propose four new techniques to perform automatic data aggregation in software DSM. Our techniques use run-time analysis of past data-fetch accesses made by a processor, to aggregate data movement for future accesses. They do not need any additional compiler support. We implemented our techniques in the TreadMarks software DSM system. We used a test suite of four applications--3D-FFT, Barnes-Hut, Ilink and Shallow. For these applications we obtained 40% to 66% reduction in message counts which resulted in 6% to 19% improvement in execution times.

[发布机构] Rice University