I am interested in the design of large, complex, networked applications and systems. My research style is engineering-oriented, prototype- and artifact-driven, and highly interdisciplinary, leading to extensive collaboration with other faculty. Our methodology depends critically on the three step process of (1) evaluating existing systems to understand their performance and capability limitations, (2) extensive simulation-based analyses to explore the design space of new solutions and architectures, and (3) implementation and measurement of the most attractive design to uncover its implementation complexities and to validate the simulation models used in the preceeding step. In a typical project, these steps are iterated two to three times.
Three years a group of faculty (Jordan, Joseph, Patterson, Shenker, Stoica, and myself) began discussing how we can leverage machine learning techniques to build distributed systems that are significantly more reliable and resilient to failures and attacks. The basic idea is to understand system building blocks spanning networks, services, and applications that can diagnose their problems and respond to them. This project is called RADS for Reliable Adaptive Distributed Systems. It has developed into some unexpected directions, emerging as one of the pre-eminent research projects on Cloud Computing. I am particularly interested in three aspects of Internet Datacenters that support the Cloud: (1) how to achieve high performance, in terms of processing speed and energy consumption, of data-intensive applications (these are the data analytics going on behind the scene at virtually all the major web sites), (2) how to collect and analyze information about the execution of applications in a complex datacenter environment of numerous machines and their topology of interconnection, and (3) how to more effectively use the network in the datacenter, which is power-intensive though relatively underutilized.
A new project, joint with David Culler, Eric Brewer, and Seth Sanders is LoCal: A Network Architecture for Localized Electrical Energy Reduction, Generation and Sharing. It investigates Information Age approaches for managing society's most limited resource: energy. The world’s electric grids are an engineering wonder of last century's physical age, each with a vast geographic reach, epitomized by a highly centralized, synchronized, and reliable distribution tree that allows electric power to be consumed without concern for its source. But rapidly changing energy demands, incorporation of non-dispatchable renewable sources, and the need to proactively manage load, have pushed this aging marvel to its limit. As the rise in greenhouse gases threatens civilization, it is time to examine how pervasive information can fundamentally change the nature of energy production, distribution and use. Taking guidance from the design principles of the dominant triumph of the cyber age, the Internet, we investigate how to design an essentially more scalable, flexible and resilient electric power infrastructure-one that encourages efficient use, integrates local generation, and manages demand through omnipresent awareness of energy availability and use over time. The crucial insight is to integrate information exchange everywhere that power is transferred. The LoCal Energy Network is a cyber overlay on the energy distribution system in its various physical manifestations, e.g., machine rooms, buildings, neighborhoods, isolated generation islands and regional grids. Pervasive information exchange will enable a more efficient scalable energy system with improved resilience and quality of delivered power. Our key contribution is to bring together (1) pervasive information about energy availability and use, (2) interactive load/supply negotiation protocols, (3) controllable loads and sources, and (4) logically packetized energy, buffered and forwarded over a physical energy network. Together these yield a system for agile, distributed, and integrated management of energy that can buffer energy on the path to reduce peak-to-average energy consumption, moderate infrastructure provisioning, and encourage power-limited design and operation. Our building block is the intelligent power switch, logically connecting sources to loads by bundling information (bits) with energy (electrons) flows.