Abstract

Why do we build data center clusters in the way that we do? I will describe some of the work we are currently doing in the CamCube project which aims to build, from the ground up, a new data center cluster architecture designed to make it easier to create the large-scale services run on these clusters. CamCube liberally borrows ideas from high performance computing, distributed systems and networking and represents a very different design point that blatantly violates many accepted norms of data center cluster architecture. The talk will motivate the design choices, including the use of a direct connect network and a new network stack. In order to demonstrate the potential of CamCube, we will describe and show results for a MapReduce-like service, that provides significantly better performance when run on CamCube compared to a traditional cluster.

Bio: For the last 11 or so years Ant Rowstron has been working as a researcher at Microsoft Research in Cambridge, UK where he is now a Principal Researcher and co-leads the Systems and Networking Group. His research interests are broad, covering the spectrum of systems, distributed systems and networking, but he is best known for his work on structured overlays and distributed hash tables. Prior to that he worked as a Research Associate and then a Senior Research Associate in Andy Hopper’s group in the CL/Engineering Department.