Prediction techniques to improve memory performance

An-Chow Lai, Purdue University

Abstract

Efficient data supply to the processor is the one of the keys to achieve high performance. However, the existing processor and memory performance gap sets the limit on how fast the data can be supplied. Unfortunately, recent trend shows that the performance gap will only increase continuously. Without improving the memory performance, architects will not likely to improve system performance significantly. In this thesis, I studied various prediction techniques which enable effective speculative data movements, such as data prefetching and data forwarding, to improve the memory performance in both uniprocessor and multiprocessor environments. There are two key elements in achieving an effective speculative data movement. First, in order to overlap the long memory access latency, each speculation has to be triggered ahead of time long enough before its corresponding memory access happens. In this thesis, I proposed a novel Last-Touch Prediction (LTP) which significantly improves timeliness in triggering speculative data movement over previous proposals. Next, one has to tell what is the target address for each speculative data movement. I proposed a Dead-Block Correlating Predictor (DBCP) to predict the address to be accessed speculatively. In the case of data prefetching, DBCP not only enables timely and accurate prefetches, but also obviates the need of using extra buffer and incurring additional prefetch hit time. In multiprocessor, it is also essential to speculatively move data to the right sharer besides timely and accurate prediction of the speculative movements. In other words, the design of a successful speculative data movement in multiprocessor environment requires a third element to tell how data are shared in the systems. In this thesis, I proposed a novel Memory Sharing Predictor (MSP) to predict the next sharer of each actively shared block. Finally, my results indicated that my proposals can effectively extract the repetitive access patterns in a program to achieve high prediction coverage and accuracy and lead speculative data movement techniques to success.

Degree

Ph.D.

Advisors

Falsafi, Purdue University.

Subject Area

Electrical engineering|Computer science

COinS