Parallel Performance Project Research Paper

Research Paper

On Effective Data Supply for Multi-Issue Processors
Jude A. Rivers, Edward S. Tam and Edward S. Davidson
Proceedings of the International Conference on Computer Design, To Appear, October 1997.

Abstract

Emerging multi-issue microprocessors require effective data supply to sustain multiple instruction processing. The data cache structure, the backbone of data supply, has been organized and managed as one large homogenous resource, offering little flexibility for selective caching. While memory latency hiding techniques and multi-ported caches are critical to effective data supply, we show in this paper that even ideal non-blocking multi-ported caches fail to be sufficient in and of themselves in supplying data. We evaluate an approach in which the first level (L1) data cache is partitioned into multiple (multi-lateral) subcaches. The data reference stream of a running program is subdivided into two classes, and each class is mapped to a specific subcache whose management policy is more suitable for the access pattern of its class. This sort of selective organization and caching retains more useful data in the L1 Cache, which translates to more cache hits, less cache-memory bus contention and overall improvement in execution time. Our simulations show that a multi-lateral L1 cache of (8+1)KB total size generally performs as well as, and in some cases better than, an ideal multiported 16KB cache structure in supplying data.