Dynamic Frequency Scaling Schemes for Heterogeneous Clusters under Quality of Service Requirements

Abstract

Nowadays, both the performance and power consumption for modern server clusters and data centers must be considered to reduce the maintenance cost for quality of service guarantees, as power dissipation affects the cost of both the power delivery subsystems and cooling facility. Considering the popularity of heterogeneous clusters, this paper proposes efficient and effective power management schemes for large scale server farms. Distinct from existing heuristic approaches, we propose dynamic frequency scaling schemes with approximation factor guarantees, compared to the optimal power management. By considering systems with discrete frequency levels on every server, our schemes can be applied for different power consumption models. Our greedy power management schemes have 1.5 or 2 approximation guarantees depending on the complexity. Our dynamic-programming approach can trade the quality, in terms of power consumption, of the resulting solutions with the time/space complexity. We provide extensive simulation results to show that the proposed schemes are effective for the minimization of the power consumption for large scale clusters.