Proactive workload consolidation for reducing energy cost over a given time horizon

Data centre energy requirements have grown massively in the last few years. One of the optimisation challenges for reducing its energy requirements is to keep servers well utilised by deciding which Virtual Machines (VMs) to migrate, where to migrate, when to migrate, and, when and which servers to switch on/off. Achieving this goal optimally requires the capability of predicting the future time-variable resource demands of VMs accurately and computing the plan for migrating VMs for efficient workload consolidation quickly. We call this Proactive Workload Consolidation Problem (PWCP). Solving PWCP as a giant monolithic problem with infinite time windows is impossible both for forecasting demands and optimal assignments of VMs to servers. We formulate PWCP in a more realistic way by defining a time window of a particular size in which the information is known more accurately and solve a - possibly infinite - sequence of optimisation problems moving forwards in time. The question is how far one is required to look ahead in terms of the number time-periods and still retain the minimum energy cost of a given horizon without violating the Service Level Agreements (SLAs). We perform investigations to understand the relationship between the number of time-periods considered in one optimisation step and migration-limits on the SLAs, energy cost, server-transition cost and migration cost. Our results suggest that looking ahead by only a few more time-periods can lead to more efficient resource provisioning over the entire horizon and consequently higher energy efficiency and close to no SLA violations.