RSRG Seminar

Recent years have seen an unprecedented growth in the number of data centers being deployed by large online service providers (OSPs). These data centers store massive amounts of data and host a variety of services. To improve performance and reliability, multiple data centers are deployed to cover large geographical regions with high-speed backbone networks interconnecting them. These backbones are often owned by the same OSP; for instance, Google, Yahoo!, and Microsoft interconnect their multiple large-scale data centers with their own private backbones. As the backbones themselves represent a substantial investment, it is highly desirable that they be used efficiently and effectively, while simultaneously respecting the characteristics of the traffic they carry.

In this talk, I will present two scalable designs for managing inter-data center traffic that lie at practical points along the spectrum between fully-distributed and fully-centralized solutions. Our designs jointly optimize rate control, routing, and link scheduling by distributing information and computation across multiple tiers of an optimization machinery. The design choices are motivated by the advantages of centralized traffic engineering and its industry adoption, as well as on the advantages of distributed rate control with low message-passing overhead. Using decomposition principles, we show that both the designs provably maximize the aggregate utility of traffic over all classes and data centers.