Cutting the Electric Bill for Internet-Scale System

Based on this observation that energy prices differ around the world — and in different ways at different times — this paper proposes redirecting traffic to utilize energy at the cheapest (sufficiently close) site. The price differences the paper speaks of exploiting are not just diurnal and seasonal ones, but hourly ones from the electricity spot market, which have substantially larger price fluctuations that cannot be definitely planned for hours in advance.

The paper uses traces of both these electricity spot prices and Akamai's workload to evaluate the potential savings from moving computation where energy is cheaper. Some practical concerns with moving workload were modeled: keeping work close to users due to latency concerns, not dramatically increasing bandwidth costs, and the realities of not having a power-proportional datacenter. Under somewhat optimistic scenarios, electricity cost reductions of more than 30% seemed achievable with this type of optimization.

The simulation evaluation was very thorough and used a real workload (which are quite hard to come by). The real worklaod used here does not, however, sound like the type one would want to prove this system by: Akamai's customers are presumably paying in large part for Akamai's closeness to the end-users, but increased distance is just about the only compromise that cannot be avoided. Surprisingly, the US electricity market seems dynanmic enough that even allowing for modest compromises (e.g. less than 1000 miles) can provide substantial benefit.