Xenmvm: Exploring Potential Performance Of Virtualized Multi-Core Systems

Zhiyuan Shao,Hai Jin,Yong Li,Jian Huang
2011-01-01
Abstract:In this paper, we propose a computing resource management system, named as XenMVM, for the virtualized multi-core platforms by Xen. It adjusts the computing resource dynamically according to the actual workload generated by the applications running inside the virtual machines to improve the resource utilization of the computer systems. Base on the knowledge of the shared L2 cache architecture of the multi-core systems, XenMVM adopts the Underlying Layout Aware Scheduling (ULAS) algorithm, which can schedule the virtual processors of the virtual machines to appropriate physical cores. The experiment data show that ULAS algorithm can improve the performance of HPC workloads by about 4.5% to 32.52%. Moreover, to improve the overall performance of overcommitted virtualized multi-core host system, where multiple virtual machines are running on, we propose two scheduling algorithms in XenMVM. Experiment data show that on processing a typical set of workloads on an 8-core platform, the proposed Domain-Based Static Priority (DBSP) algorithm can reduce the turn-over time of the whole system by about 20.45% compared with the proposed First Come First Serve (FCFS) algorithm.
What problem does this paper attempt to address?