Adaptive Power and Resource Management Techniques for Multi-threaded Workloads

Abstract

As today's computing trends are moving towards the cloud, meeting the increasing computational demand while minimizing the energy costs in data centers has become essential. This work introduces two adaptive techniques to reduce the energy consumption of the computing clusters through power and resource management on multi-core processors. We first present a novel power capping technique to constrain the power consumption of computing nodes. Our technique combines Dynamic Voltage-Frequency Scaling (DVFS) and thread allocation on multi-core systems. By utilizing machine learning techniques, our power capping method is able to meet the power budgets 82% of the time without requiring any power measurement device and reduces the energy consumption by 51.6% on average in comparison to the state-of-the-art techniques. We then introduce an autonomous resource management technique for consolidated multi-threaded workloads running on multi-core servers. Our technique first classifies applications according to their energy efficiency measure, then proportionally allocates resources for co-scheduled applications to improve the energy efficiency. The proposed technique improves the energy efficiency by 17% in comparison to state-of-the-art co-scheduling policies.