Distributed temperature-aware resource management in virtualized data center

Abstract

While resource consolidation enables energy efficiency in virtualized data centers, it results in increased power density and causes excessive heat generation. To prevent servers from overheating and avoid potential damage and/or service outages, data centers need to incorporate temperature awareness in resource provisioning decisions. Moreover, data centers are subject to various peak power constraints (such as peak server power) that have to be satisfied at all times for reliability concerns. In this paper, we propose a novel resource management algorithm, called DREAM (Distributed REsource mAnagement with teMperature constraint), to optimally control the server capacity provisioning (via power adjustment), virtual machine (VM) CPU allocation and load distribution for minimizing the data center power consumption while satisfying the Quality of Service (QoS), IT peak power and maximum server temperature constraints. By using DREAM, each server can autonomously adjust its discrete processing speed (and hence, power consumption, too), and optimally decide the VM CPU allocation as well as amount of workloads to process in the hosted VMs, in order to minimize the total power consumption which incorporates both server power and cooling power. We formally prove that DREAM can yield the minimum power with an arbitrarily high probability while satisfying the peak power and server temperature constraints. To complement the analysis, we perform a simulation study and show that DREAM can significantly reduce the power consumption compared to the optimal temperature-unaware algorithm (by up to 33%) and equal load distribution (by up to 86%).