Operating policies for energy efficient large scale computing

Abstract

Energy costs now dominate IT infrastructure total cost of ownership, with datacentre operators predicted to spend more on energy than hardware infrastructure in the next five years. With Western European datacentre power consumption estimated at 56 TWh/year in 2007 and projected to double by 2020, improvements in energy efficiency of IT operations is imperative. The issue is further compounded by social and political factors and strict environmental legislation governing organisations. One such example of large IT systems includes high-throughput cycle stealing distributed systems such as HTCondor and BOINC, which allow organisations to leverage spare capacity on existing infrastructure to undertake valuable computation. As a consequence of increased scrutiny of the energy impact of these systems, aggressive power management policies are often employed to reduce the energy impact of institutional clusters, but in doing so these policies severely restrict the computational resources available for high-throughput systems. These policies are often configured to quickly transition servers and end-user cluster machines into low power states after only short idle periods, further compounding the issue of reliability. In this thesis, we evaluate operating policies for energy efficiency in large-scale computing environments by means of trace-driven discrete event simulation, leveraging real-world workload traces collected within Newcastle University. The major contributions of this thesis are as follows: i) Evaluation of novel energy efficient management policies for a decentralised peer-to-peer (P2P) BitTorrent environment. ii) Introduce a novel simulation environment for the evaluation of energy efficiency of large scale high-throughput computing systems, and propose a generalisable model of energy consumption in high-throughput computing systems. iii iii) Proposal and evaluation of resource allocation strategies for energy consumption in high-throughput computing systems for a real workload. iv) Proposal and evaluation for a realworkload ofmechanisms to reduce wasted task execution within high-throughput computing systems to reduce energy consumption. v) Evaluation of the impact of fault tolerance mechanisms on energy consumption.