Fast demand response with datacenter loads: a green dimension of big data

When resources are limited, most people benefit from “playing nice.” The most effective kind of sharing is about more than politeness and good manners. It’s about relationships and trust. Those things take a lot of effort and understanding – especially between people with opposing goals – but they’re worth it in the end. The idea of datacenter demand response is to design a mutually-beneficial relationship between an unpredictable electricity user (load) and a cautious electricity supplier (utility). The challenge is a difficult one, but the benefits to both participants are worth it in the end.

Demand response is a way for large electric loads to “play nice” both with utilities and with other loads that can’t control their own power consumption. Sharing strategies are beneficial to both loads and utilities because electric power is a fundamentally limited resource. This isn’t just because most power plants run off of limited energy sources like coal and natural gas. At any given instant, loads can use no more power than is being generated at that moment. And, to avoid power surges, utilities have to pay money to “dump” any excess power that they generate. It’s impossible to predict exactly how much power will be needed at any given moment, and renewable energy suppliers like wind and solar make it difficult for utilities to precisely control the amount of power they generate. If some loads help utilities out by temporarily raising or lowering their power consumption, the delicate balance between generation and load is easier to maintain and everyone receives cheaper, more reliable power.

Datacenter operators have become so good at providing transparent access to Internet applications, scientific computing, business analytics and other services, that it’s easy to forget that “the cloud” is a collection of physical machines that need regular maintenance and lots of electrical power. Datacenters’ large, quickly-changing, and unpredictable electric load currently makes it more difficult for utilities to pro- vide reliable power. However, some of the very features that make datacenters such difficult customers can actually let datacenters provide demand response services that other controllable loads can’t. In particular, datacenters are uniquely well-suited to provide realtime demand response services because of their ability to quickly change their power consumption across a wide range of power levels. Other types of demand response (peak capping, emergency service) can help with long-term reliability planning. However, realtime demand response is especially valuable in the fast-changing power grid of today, because it can respond quickly and continuously to unpredictable loads and renewable generation.

The benefit both to utilities and datacenters is very much worth the effort of making datacenter demand response a reality, and this thesis answers several questions related to that goal: As a way of showing that power shaping is actually practical, it presents experimental video transcoding service that also provides a high quality fast power shaping service to the electric grid. Next, we explore how power shaping affects the performance of two common datacenter workloads: Read-optimized queries (used to quickly explore large datasets which do not change very often), and update- aware indexing (used to speed up access to large, frequently-changing datasets such as Twitter trending tweets). In these contexts, we show that certain performance optimizations can also have a positive impact on power requirements. This reduction in required power consumption gives clusters serving these workloads the flexibility they need to participate in load shaping programs. We finish up by mentioning some preliminary ideas on designing power controllers that are robust to the complex and quickly-changing environments of real-world datacenters.