Real-Time Power Cycling in Video on Demand Data Centres using Online Bayesian Prediction
Vicent Sanz Marco, Zheng Wang and Barry Porter
Lancaster University, Lancaster University, Lancaster University

Energy usage in data centres continues to be a major and growing concern as an increasing number of everyday services fundamentally depend on these facilities. Research in this area has examined topics including power smoothing using batteries and deep learning to control cooling systems, in addition to optimisation techniques employed in the software running inside data centres. We explore the technique of server power cycling in which unused servers are automatically switched off when not needed, and later switched on when demand increases. We specifically examine this approach in video on demand systems which tend to use dedicated servers to ensure predictable quality of service. We present a novel distributed architecture, media distribution approach, and a real-time prediction model to determine when to most effectively switch servers on and off, and demonstrate with experimental evaluation that this approach can save up to 31% of server energy in a cluster. Our evaluation is conducted on typical rack mount servers in a production data centre and uses recent real-world workload trace from BBC iPlayer, an extremely popular video on demand service in the UK.