PATPro: Power Aware Thin Provisioning of Resources in Virtualized Servers
Congfeng Jiang, Yinghui Zhao · Applied Mathematics & Information Sciences · 2013
In traditional virtualization systems Virtual Machines (VMs) are usually over-provisioned for guarantying peak performance of hosting applications and thus waste a lot of computing resources. Although virtual machine consolidation can save power consumptions, it also increases the power intensity in a rack and makes the servers more prone to failures. It also increases the probability of Service Level Agreements (SLAs) violations under constrained power budget. Due to the autonomy, workload dynamics and performance requirements of Virtual Machines, it is nontrivial to tradeoff between the power and performance in virtualized servers. In this paper, we proposed a power aware thin provisioning approach for resource allocation, named PATPro, to coordinate and tradeoff the power and performance among multiple virtual machines. In order to provide enough information about power characteristics of individual VM and applications, we use agent for real time power collection and workload characterization. We evaluated our algorithm in a real virtualization environment with 8 VMs. The results show that the proposed approach can save 5.41% to 6.33% power consumption to provide the same performance in applications including computing-intensive, I/O intensive and hybrid workloads such as web servers and database operations. The results also show the potential of our approach to be used in real virtualized environments.