Multilayer cache partitioning for multiprogram workloads

摘要: We present a fully-automated, model based, multilayer cache partitioning scheme for multiprogram workloads running on multicore machines. As opposed to prior efforts, this partitions shared caches at multiple layers simultaneously in coordinated fashion. This tries achieve two objectives. First, it satisfy the specified quality of service (QoS) values all applications by hierarchy across them, and second, distributes remaining excess capacity (if any) such that global performance metric is maximized. Our experimental analysis shows proposed generates, average, 33.1% improvement (on weighted speedup metric) over next best-performing very successful satisfying QoS requirements applications. Also, we show each layer isolation cannot generate benefits obtained through our scheme. In addition, observed difference between an optimal (that derives best dynamic partitions) was less than 15% tested 6.6% average.