Energy-aware workload management models for operation cost reduction in data centers

In the last century, the costs of powering datacenters have increased so quickly, that datacenter power bills now dwarf the IT hardware bills. Many large infrastructure programs have been developed in the past few years to reduce the energy consumption of datacenters, especially with respect to cooling requirements. Although these methods are effective in lowering the operation costs they do require large upfront investments. It is therefore not surprising that some datacenters have been unable to utilize the above means and as a result are still struggling with high energy bills. In this work we present a cheap addition to or an alternative to such investments as we propose the use of intelligent, energy efficient, system allocation mechanisms in place of current packaged system schedulers available in modern hardware infrastructure cutting server power costs by 40%. We pursue both the quest for (1) understanding the energy costs generated in operation as well has how to utilize this information to (2) allocate computing tasks efficiently in a cost minimizing optimization approach. We were able to underline the energy savings potential of our models compared to current state-of-the-art schedulers. However, since this allocation problem is complex (NP-hard) we investigated various model approximations in a trade-off between computational complexity and allocative efficiency. As a part of this investigation, we evaluate how changes in system configurations impact the goodness of our results in a full factorial parametric evaluation.

► We model the dependency between server utilization and energy consumption. ► We formulate five workload management decision models with varying complexity. ► We investigate the tradeoff between complexity and solution quality. ► Using non-linear instead of linear functions does not increase efficiency as much. ► Using energy-aware schedulers saves up to 40% energy costs vs. standard schedulers.