Edinburgh Research Explorer

LLC Dead Block Prediction Considered Not Useful

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Related Edinburgh Organisations

Access status

Open

Documents

Original languageEnglish
Title of host publication13th Workshop on Duplicating, Deconstructing and Debunking (WDDD)
Number of pages10
StateAccepted/In press - 14 May 2016

Abstract

Dead block predictors (DBPs) improve cache efficiency by identifying blocks that have exhausted their useful lifetime in the cache and prioritizing them for eviction regardless of how much reuse these blocks have experienced. In
doing so, DBPs go beyond insertion policies that target only those blocks that have no cache re-use. But how much value do DBPs add over insertion policies?
This work examines the opportunity for DBP at the LLC and concludes that its value-add is low over a state-of-the-art insertion policy. Our key result is that LLC evictions are dominated by blocks having no reuse. Even with the op-
timal replacement policy that maximizes cache reuse through perfect knowledge of the future, an average of 78% of LLC evictions have not experienced any hits. For the remaining evictions that experience LLC reuse, only a fraction is predicted by a state-of-the-art DBP scheme, and the accuracy of these predictions is low. Our first-order limit study shows that a likely coverage ceiling for DBP over the best performing insertion policy is just 6.9% of all LLC evictions. Based on these findings, we argue that the higher complexity of DBP is not justified.

Download statistics

No data available

ID: 25451536