of Computer Science Colloquium
Thursday, November 7, 2002, 4:15pm
Upson Hall B17
Prediction-Based Page Prefetching to Improve the Performance
of Shared Virtual Memory Systems
Software distributed shared memory (SDSM) systems traditionally exhibit poor performance on many applications. Techniques such as relaxed-consistency models and multiple-writers protocols improve the performance of SDSM systems significantly, but their performance still lags that of hardware shared memory implementations. This talk will describe Delphi, a system that borrows techniques from microarchitectural research on value prediction and applies them to software distributed shared memory. We run a small software predictor on each node in the Delphi system to predict which virtual pages will be needed in the future. We use these predictions to prefetch pages in order to reduce the number of accesses to invalid data and thereby reduce expensive network accesses. Experimental results show that Delphi is able to reduce the number of read misses to virtual pages significantly on a set of well-known scientific benchmarks, with a concomitant decrease in execution time.