Evaluating the viability of remote renewable energy in datacentre computing

We investigate the feasibility of loosely-coupled distributed datacentre architectures colocated with and powered by renewable energy sources and interconnected using high-bandwidth low-latency data links. The design of these architectures advocates (i) variable machine availability: the number of machines accessible at a particular site at any given time is proportional to the amount of renewable energy available and (ii) workload deferment and migration: if there is insufficient site capacity, workloads are either halted or shifted to transient energy-rich locations. While these architectures are attractive from an environmental perspective, their feasibility depends on (i) the requirement for additional hardware, (ii) the associated service interruptions, and (iii) the data and energy overhead of workload migration. In this work we attempt to broadly quantify these overheads. We define a model of the basic design of the architecture incorporating the energy consumption of machines in the datacentre and the network. We further correlate this energy consumption with renewable energy available at different locations around the globe. Given this model we present two simulation-driven case studies based on data from Google and Facebook production clusters. Generally we provide insights on the trade-offs associated with this off-grid architecture. For example, we show that an optimised configuration consisting of ten distributed datacentres results in a 2% increase in job completion time at the cost of a 50% increase in the number of machines required.