RDPart: A reuse-based OS-level cache-partitioning policy for fairness optimization in cloud data centers

Abstract

In cloud data centers, the colocation of multiple services and applications on the same physical server is crucial for maximizing resource utilization and reducing utility costs. Unfortunately, contention for shared resources across cores, such as the Last-Level Cache (LLC), may lead to severe interference among applications. This complicates quality-of-service (QoS) enforcement and causes performance disparities among applications, ultimately degrading user experience and system fairness. To address these issues, this paper proposes RDPart, an OS-level Reuse-Driven LLC Partitioning policy designed to improve fairness while preserving the QoS of cloud workloads. In contrast to existing fair LLC-partitioning techniques, RDPart eliminates the need for online performance profiling under varying LLC allocations, thereby avoiding undesired QoS violations typically stemming from such invasive exploratory approaches. Furthermore, RDPart adopts a black-box design, making it well-suited for public cloud environments where real-time QoS feedback from applications is unavailable. We implement RDPart in the Linux kernel and evaluate its effectiveness across a diverse set of workloads, including cloud services and HPC/scientific applications. The results reveal that RDPart achieves a 32.1% average fairness improvement with respect to a state-of-the-art black-box cache-partitioning proposal, while keeping QoS violations consistently low across workloads.