Cutting the Electric Bill for Internet Scale Systems

Summary

This paper describes how electricity prices vary with time and geographic location, and how data centers can take advantage of this information to save in electricity costs.

They rely on 2 key observations:

1. electricity prices vary
2. large distributed systems already incorporate request routing and replication

Using price information, the goal is to have a system route client requests to clusters such that total electricity cost is minimized. This also depends on the energy elasticity: the extent to which a cluster's electricity costs depend on the load on that cluster.

By simulating 29 different locations in the US with historical electricy prices, the authors get results that suggest huge cost savings for data centers, including:

• existing systems can see energy cost savings by at least 2%
• savings rapidly increase with energy elasticity
• allowing client-server distances to increase leads to increased savings

Criticism & Questions

This was a very interesting paper to read. I liked how the paper problem was rooted in a practical problem with a very practical solution. This paper was recently published, but I would be interested in knowing if any data centers have tried implementing this solution and what kinds of results they're getting.