RADAR: Runtime-Assisted Dead Region Management for Last-Level Caches
Last-level caches bridge the speed gap between processors and the off-chip memory hierarchy and reduce energy per access. Unfortunately, last-level caches are poorly utilized because of the relatively large occurrence of dead blocks; blocks that are not accessed before becoming evicted. In particular, dead-block prediction is
challenged by unpredictable scheduling decisions made in run-time systems supporting task parallel programming models.
This paper presents RADAR, a hybrid hardware/software dead-
block management scheme, that can accurately predict dead blocks. It does so by inferring dead blocks from data-flow information about adress regions through functionality built into the run-time system and uses hardware support to evict dead blocks belonging to such regions. RADAR utilizes semantic information about address regions provided by programmers, past access patterns, and future reuse information available to the runtime system to accurately
predict dead regions during execution. Our evaluation shows that RADAR outperforms state of the art dead block prediction schemes that rely only on the past eviction patterns to predict dead blocks.
last level cache
task-based programming models
dead block prediction