This program undertakes a broad research agenda centered around the design and analysis of ``Flow-based Networks''. A flow is a collection of packets that belong to the same ``transaction'', such as a datagram, an ftp transfer, or a web download. It is the fundamental unit of data that a user cares about. Current packet-switched networks, like the Internet and Gigabit Ethernet, are designed to process packets; they are unaware of the flow to which a packet belongs. This is because flow-recognition is widely considered to be too expensive to implement. However, a switch or a router's ability to recognize flows can lead to a marked improvement in its performance, to a better use of its resources, and to much more secure networks.
The first major aim of this program is to design novel algorithms and data structures for high-speed, "flow-aware" networks. Such algorithms could heavily influence the design of commercial switches and routers. A second major thrust concerns the development of flow-level models of networks: models which capture the impact of packet-level decisions on flow-level bandwidth allocation and flow processing times. An important component of the modeling work is the unification and generalization of two researce enterprises: Stochastic Network Theory, and Large Random Networks. The former studies the performance of a, typically non-random, queueing network subject to "random inputs". The latter concerns the study of "random networks", usually subject to deterministic inputs. A successful outcome of these efforts can help answer questions such as the throughput and flow delay of a particular bandwidth allocation scheme, and the effect of routing topology on end-to-end performance. In other words, the modeling effort aims to develop a realistic, simple and usable class of models for network flows.
