I/O-Conscious And Prediction-Enabled Virtual Machines Scheduling

Jinlei Jiang,Xun Zhao,Yongwei Wu,Weimin Zheng

DOI: https://doi.org/10.1109/CIT.2016.92

2016-01-01

Abstract:Virtual machine (VM) technology plays an important role in modern cluster and cloud computing environments due to its advantages in application isolation, resource partition and load consolidation. Since a large number of VMs usually run simultaneously in these environments, VMs scheduling is key to achieve resources efficiency. Unfortunately, most existing scheduling strategies are complicated, with unacceptable overhead in a large-scale virtualization environment. Further more, they tend to produce biased results without considering the I/O characteristics of VMs and physical machines (PMs).To deal with the issue, this paper proposes a new strategy for scheduling VMs over physical clusters. Taking into consideration the I/O characteristics, the proposed strategy can obtain better scheduling results while with lower overhead. Moreover, a prediction algorithm is suggested, which, when combined with the proposed strategy, can further improve the scheduling results. The proposed strategy and prediction algorithm have been implemented and evaluated with the SPEC CPU 2006 and Netperf benchmarks. The experimental results show that the average completion time of SPEC reduces by about 23%, whereas the standard deviation of Netperf average bandwidth decreases from 10.9 to 4.2.