VM Co-scheduling: Approximation of Optimal Co-scheduling in Data Center

Multi-resource sharings among virtual machines (VM)are very important in data center, but also bring contentions for cache, disk and network in host server. Recent studies focus primarily on locally optimizing resource sharings in a single server, or global optimization of mitigating cache and memory contentions. But it remains an open question how to efficiently find optimal co-schedules for cache, disk and network contentions simultaneously. Solving these contentions is critical for computing-intensive, memory-intensive, data-intensive and network-intentive applications. This paper is the first to present a comprehensive solution to reduce multiple resource contentions in terms of cache, disk and network contentions in data center. We propose two estimation models, MPain and MVector Distance, to predict co-run performance degradation based on the VM solo running information. Experiment results on real servers reveal that the MVector Distance is more effective than up-to-date task activity vector model in EuroSys´10, reducing the average performace degradation from 8.545% to near zero in exhaustive search cases. We develop both proactive and runtime reactive approximation algorithms, among which, the simple balance algorithm is a three-dimensional coupling algorithm, produces near optimal schedules and shows good scalability. This study offers some techniques which could be directly used in proactive and runtime reative VM co-schedulings.