Applications of Power Demand Routing in Distributed Systems

Опубликовано 17 августа 2016, 20:56

This talk will cover our work exploring the role of software in modulating the flow of electric power within Internet systems, to reduce operating costs and environmental impact. Large Internet services are implemented using hundreds of thousands of servers, consuming enough electricity to power thousands of homes. Unsurprisingly, there is a desire to reduce both electricity cost and consumption. The majority of the talk will focus on the concept of 'Power Demand Routing' and its role in reducing energy costs, without reducing consumption. Within existing web services, each client request requires a meaningful amount of marginal energy at the server. Thus, by rerouting requests from a server at one geographic location to another, we can spatially shift the systemΓÇÖs marginal power consumption at Internet speeds. One application of this is to track geographic variations in the service cost (e.g., variations caused by price fluctuations in spot electricity markets) within a replicated system, and to adapt request routing policy to take advantage of meaningful transient cost-differentials between replicas. This skews client load and pushes server power demand into the least expensive regions. Our analysis quantifies the potential reduction in system operating costs, taking performance and bandwidth costs into account. Another application stems from the fact that not all watts are created equal. In power pools, like the grid, that aggregate electricity from diverse providers, the environmental impact per watt varies in time and can be uncorrelated at different locations, depending upon various factors, such as what local generating assets are active (e.g., wind vs. natural gas). Rather than attempting to minimize the dollar cost of the energy consumed, a service operator may instead choose to use power demand routing with an environmental impact cost function. More recently we have been looking closely at how modulating the flow of power in smaller-scale distributed systems can reduce energy consumption. In particular we have designed an adaptive data-block layout strategy for Web-service storage clusters that takes advantage of non-uniform block popularity and the heterogeneous power consumption curves of different classes of processor hardware to design a low-power storage cluster. This talk covers joint work with John Guttag (MIT), Hari Balakrishnan (MIT), Bruce Maggs (CMU/Duke/Akamai) and Rick Weber (Akamai).