Design and Evaluation of a Distributed Cache Architecture with Prediction

Thomas Alexander and Gershon Kedem

abstract

We propose a secondary cache architecture that combines a predictive fetch strategy with a distributed cache to build a high performance memory system. The cache is partitioned into smaller units and distributed evenly in the main memory space. The architecture offers high bandwidth between the cache and the DRAM memory. A hardware prediction scheme is used to prefetch data into the cache and hide the high DRAM latency. The prediction scheme does not rely on any predetermined data access patterns and is completely transparent to the user. Simulation of our architecture on a set of benchmark programs showed a 40%-90% improvement in the effective memory access time when compared to traditional caching.

On-line project information

Introductory Slides

Thomas Alexander and Gershon Kedem, "Distributed Prefetch-buffer/Cache Design for High Performance Memory System." Procedings HPCA-2. pp. 254-263, Feb. 1996. compressed postscript format and PDF format.

HPCA-2 presentation slides.

Thomas Alexander, "A Distributed Predictive Cache for High Performance Computer System", PhD Thesis, Duke University, 1995. compressed postscript format and PDF format. (big)