Fluid Models for Large, Heterogeneous
Networkss
Summary
Researchers from the University of Massachusetts propose to perform fundamental research on developing novel fluid-based methodologies that will enable the solution of large networks handling large numbers of responsive flows (e.g., TCP) and non-responsive flows (e.g., video). These methodologies will provide for the rapid and efficient solutions of these networks to determine the throughputs, average delays, and loss rates of individual and aggregate flows. In addition to its use in analyzing large networks of TCP and UDP flows, our methodology will also permit researchers to evaluate the performance of new transport protocols (e.g., reliable multicast) in isolation as well to study their impact on legacy TCP flows. The novel features of the effort include:
- Fluid-based approximation of networks supporting long-lived responsive flows. We will develop computationally efficient and accurate techniques for solving a network of routers, implementing active queue management policies (e.g., RED), and supporting large numbers of long-lived responsive flows and non-responsive flows. These techniques can be used to estimate flow throughputs, average delays, and loss rates in large scale networks. (By ``large scale,'' we mean large in terms of the number of routers and the number of flows).
- Fluid-based transient analysis of networks supporting responsive flows. We will develop computationally efficient and accurate techniques for approximating the transient behavior of responsive and non-responsive flows in a network. Such transient techniques are particularly useful in networks that support short-lived responsive flows along with long-lived flows.
- FAN, a Fluid-based Analysis of Networks tool. The techniques we develop during the course of our research will be incorporated in an easy to use tool for network engineers. The tool will be publicly available and will be a key element of our technology transfer.
- Application to active queue management. Using the above methodologies, we will develop computationally efficient techniques to evaluate the performance of new active queue management policies in large-scale networks. To date, evaluation of such new policies has been confined to small-sized networks, due to our inability to solve large scale network models with active queue management. Our research will remove this roadblock.
- Application to differentiated services architecture. Our methodologies will also be applicable to analyzing responsive flows within the emerging differentiated services framework. As with active queue management, evaluation of Diff-Serv architectures has been confined to small-sized networks; our research will allow large-scale Diff-Serv networks to be studied. Examining Diff-Serv in a large-scale setting is crucial since the issue of scale is a prime motivation behind Diff-Serv.
Comments: towsley@cs.umass.edu