Locpart: A Latency Optimized Cache Partitioning for Cloud Data Centers

As continual evolution of virtualization technology and cloud computing paradigm, modern data centers consolidate various application to run with less servers. However, co-locating applications is challenging, especially for mixed workloads of both latency-critical and batch programs. It is required to address the unpredicted performance and quality-of-service (QoS) due to contention for the shared last-level caches (SLLC). In this paper, we present a novel latency optimized cache partitioning policy, Locpart, which requires neither hardware modification nor complicated software design. Locpart prioritizes latency-critical services over batch workloads on generating cache partitioning schemes with conservative allocation to gurantee QoS. At the same time, Locpart seeks to balance the rest cache ways among the workloads to achieve a maximum global performance. Evaluation with workloads mixes of emerging latency-critical as well as conventional batch programs show that Locpart guarantees the QoS and improves the global system performance in parallel significantly.