A Fast and Flexible Transport Architecture for High Speed Networks

Modern datacenter networks and private wide area networks underpin cloud computing. Today, due to the lack of flexibility, operators end up over provisioning their networks significantly to avoid performance bottlenecks. This project will make these networks more efficient and more tuned to the application's needs. Specificallyt, this investigation explores a weighted transport abstraction as a flexible and robust substrate for systems that optimize a network's bandwidth allocation. The proposed research revisits a classic theoretical framework in this space, Network Utility Maximization (NUM), and develops a novel and practical distributed algorithm for NUM that is significantly faster than prior approaches, and is thus applicable to modern high speed networks such as datacenter fabrics. This project seeks to design and build a flexible transport architecture.

The proposed architecture has two main technical components:

(1) Weighted Transport: Instead of using flow rates to control the bandwidth allocation, this research develops a transport based on weights. To tune the bandwidth allocation, flows adapt a weight field in their packet headers; each link then divides its bandwidth among contending flows in proportion to their weights.

(2) Fast Utility Maximization: This project leverages the weighted transport to design a network fabric that can be dynamically tuned for different bandwidth allocation objectives such as minimizing flow/coflow completion time, or service-level fairness.

The education plan includes the incorporation of this research's findings into the undergraduate and graduate curricula and offers an opportunity to take a "top-down" approach to teaching transport architectures with a focus on key bandwidth allocation objectives and how they affect real applications. The course material will be made widely available through MIT OpenCourseWare and on the MITx MOOC.