Abstract

Cloud computing technology offers on-demand network access to shared pools of configurable computing resources, such as virtual servers, applications and software services. This paradigm delivers the economy of scale to users via large datacenters, fast and flexible provisioning of resources, and the freedom from long-term investments in equipment. Despite the potential advantages, a new set of challenges must be overcome before cloud providers can see a return on their investments; namely, dealing with the uncertainty involved in operating cloud systems, both in terms of supply (time-varying energy prices and availability) and demand for computing resources.

One way to reduce operation costs is by exploiting the spatiotemporal variation of electricity prices. I start by showing how techniques from online algorithms can be employed to develop algorithms that move computation to datacenters in which energy is available at a cheaper price. The algorithm we design performs remarkably well on empirical electricity-prices and job-demand data. In the second part of the talk, I develop a general queuing-based model for the study of cloud computing economics, incorporating stochastic demands and diversity in the application types and user preferences. Despite the high level of heterogeneity, I show that a simple per-unit price (currently employed by most cloud platforms) suffices to induce social efficiency. I will conclude with a discussion of ongoing work on combining energy-aware scheduling and mechanism design for optimizing profits.