Last level cache partitioning via multiverse thread classification

Authors

BURAK SEZİN OVANT
İSA AHMET GÜNEY
MUHAMMED EMİN SAVAŞ
GÜRHAN KÜÇÜK

Abstract

Last level caches (LLCs) are part of the last line of defense against the famous memory wall problem. Today, almost all processors utilize a LLC for the same reason. This study extends our previous work, which proposed a cache-partitioning mechanism using thread classification. Here, we propose three enhancements to the existing system: 1) an adaptive traffic threshold mechanism for more portable classification hardware, 2) a new method for classifying way-hungry threads, and finally, 3) a thorough comparison of two design alternatives. Compared to the original way- partitioning mechanism, we show that the proposed mechanism's performance improved by around 6%, on average.

DOI

10.3906/elk-1610-336

Keywords

Cache partitioning, classification, multiverse

First Page

220

Last Page

233

Recommended Citation

OVANT, B. S, GÜNEY, İ. A, SAVAŞ, M. E, & KÜÇÜK, G (2018). Last level cache partitioning via multiverse thread classification. Turkish Journal of Electrical Engineering and Computer Sciences 26 (1): 220-233. https://doi.org/10.3906/elk-1610-336