NeTS: Small: Hybrid Switching in Data Center Networks: Systems-driven Modeling and Principled Algorithms

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

Data analytics are increasingly an engine of technical and economic growth. They typically involve distributed computation across many interconnected servers in large data centers. The increasing number of applications and the rapidly growing scale of datacenters are driving a strong need for fast, dynamic and economic changes in the interconnection patterns among datacenter servers. To address this, this project will make use of optical circuit switches, which can handle large data traffic flows with sparse connection patterns. It will develop scheduling algorithms to manage when to connect servers via these switches, taking into account practical constraints associated with the switching hardware as well as practical systems constraints. Addressing this challenge will enable more efficient use of data center resources for data analytics and other computationally intensive tasks. This project addresses the problem of fast dynamic switching for adaptation to datacenter traffic patterns. In particular it investigates the issue of balancing the overhead of reconfiguring circuit switches with the capability of dynamically meeting traffic demand requirements. The project will design scalable scheduling algorithms for hybrid switches (combination of optical circuit switches and electronic packet switches) with direct applications to modern data centers, addressing a wide range of practical requirements. These goals are addressed via a two-step approach: (a) Systems modeling, i.e. deriving canonical models of the optimization problems underlying the constraints of the hardware (speed, connectivity constraints and switching costs) as well as practical system requirements (such as throughput and tail latency); (b) Principled algorithms: deriving scalable scheduling solutions with formal optimality guarantees by conducting a first-principles study of the underlying mixed continuous/combinatorial optimization problems. The output of this project is expected to impact both practical networking designs of modern data centers as well as the development of next generation optical switches with high connectivity and small switching costs.