ACAM: Application Aware Adaptive Cache Management for Shared LLC

Abstract

Modern Chip Multiprocessors (CMPs) are typically multicore systems with shared last level cache (LLC). Effective utilization of the shared cache resource can be a challenge when the demands of competing applications conflict with each other. At times, in order to accommodate new data required by one application, the other application’s useful data may get evicted. Such negative interference results into increase in memory miss and degrades system’s performance. Hence, a technique is required which optimally manages the LLC even in the presence of such conflicting demands.

Various LLC management techniques have been proposed to efficiently manage shared caches. The state-of-the-art replacement policies like Static Re-reference Interval Prediction (SRRIP) and Application Aware Behavior Re-reference Interval Prediction (ABRip) evict a cache block based on their re-usability in the near future. SRRIP makes the replacement decisions per block basis whereas ABRip also considers the cache behavior of an application to minimize conflicting data demands. Hence, ABRip outperforms SRRIP for workload mixes where one application is cache friendly, and the other one is streaming. However, ABRip does not perform well when the workload mix is Cache friendly-Cache friendly. We propose Application Aware Adaptive Cache Management policy that adapts to both types of workload mixes. The proposed replacement policy reduces LLC misses per kilo instruction (mpki) up to 22.74% and 12.7% compared to SRRIP and ABRip respectively on a CMP system running SPEC CPU2006 workloads. Our policy effectively utilizes the shared LLC and outperforms both SRRIP and ABRip with performance gains of up to 10.12% and 9.36% respectively.