Topology‐aware Virtual Resource Management for Heterogeneous Multicore Systems
Jianmin Qiana, Jian Lib and Ruhui Mac
Shanghai Key Laboratory of Scalable Computing and Systems, Shanghai Jiao Tong University, China
ahacker qian@sjtu.edu.cn
bli-jian@sjtu.edu.cn
cruhuima@sjtu.edu.cn
ABSTRACT
Virtualization technology consolidates multiple independent workloads on a single physical server with virtual resource management, which can result in a significant utilization improvement and energy saving. However, the management of virtual resource is becoming more and more challenging, due to the lack of accurate performance prediction model for the diverse applications’ irregular resource access behaviors as well as the complicated Non‐Uniform Memory Access (NUMA) server architecture. These challenges drastically affect the overall consolidation performance. This paper proposes vTRMS, a runtime Topology‐aware virtual Resource Management Scheme for heterogeneous NUMA multicore systems. vTRMS can improve application performance based on the comprehensive online monitor of the application resource access behaviors as well as an accurate and platform‐independent detected NUMA topology metric. Experiment results show that, compared with state‐ofart approach, vTRMS can bring up an average throughput improvement of 28.3% and 36.2% on Intel and AMD NUMA machines respectively, when consolidating 32‐VMs. At the same time, vTRMS only incurs a runtime overhead no more than 5%.
