Multilayer cache partitioning for multiprogram workloads

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