CDLP: A core distributing policy based on logic partitioning

Abstract

With multi-core processors being increasingly employed by servers in cloud/data centers, how to take advantage of multi-core has been a key issue to improve the system performance. Data reuse in shared cache is an attractive approach to exploit multi-core, and therefore researchers present a number of solutions from different levels such as compiler and operating system. However, because of the lack of cooperation in two-level scheduling framework in virtualized system, the scheduling and distributing policy employed by VCPU (Virtual CPU) scheduler often counteract the endeavor of the guest operating in pursuing data reuse. Aiming the problem, we present CDLP: a novel dynamic core distributing policy based on logic partitioning, and we present a set of algorithms including core distribution, load balancing, core redistribution, core reclaiming. In order to verify the effectiveness and correctness of CDLP, we implement a prototype based on Xen platform, the experimental results demonstrate that CDLP can effectively realize data reuse and fully utilize processing resource.