Last-Touch Correlated Data Streaming

Tuesday December 7, 2004
Hamerschlag Hall D-210
4:00 pm

Mike Ferdman
Carnegie Mellon University

Effective data prefetching requires accurate mechanisms to predict both which cache blocks to prefetch and when to prefetch them. Recent research advocates using last-touch predictors to both identify the last access to a cache block prior to eviction and prefetch a replacement. Unfortunately, current last-touch predictors and prefetchers are impractical because they store stand-alone prediction signatures in on-chip associative structures and require storage proportional to a program's memory footprint.

In this talk, I will present two last-touch predictors. First, I will review the Dead-Block Correlating Prefetcher, a previously proposed last-touch predictor that requires very large on-chip storage to be effective. Second, I will introduce Last-Touch Correlated Data Streaming (LT-CORDS), a practical design for data streaming that leverages the benefits of last-touch predictors while offloading the large predictor signature storage to off-chip DRAM.

Mike is a first year PhD student in the Computer Architecture Laboratory at Carnegie Mellon, where he is advised by Prof. Babak Falsafi. His time has primarily been devoted to researching DBCP and constructing LT-CORDS. His research interests include processors and anything that relates to them.