Stability and convergence of TCP-like congestion controllers in a many-flows regime

With the rapid growth of Internet, parameter design and analysis for large-scale networks has become a topic of active interest. Since simulation of such large scale systems is not easy, deterministic fluid models have been widely used for both qualitative understanding of the behavior, as well as parameter design for such networks. In this paper, we first study a deterministic fluid model for Internet congestion control when there are multiple TCP-like flows present. We provide conditions under which such a system is globally asymptotically stable in the presence of feedback delay. We then study the corresponding system with the addition of web mice and other non-responsive flows modeled as stochastic disturbances. We show that, when there are a large number of flows, choosing parameters based on the global stability criterion for the deterministic system (with the noise replaced by its mean value) ensures global stability for the stochastic system as well. Numerical examples and simulation results with some popular active queue management mechanisms validate the parameter choices from analysis. The results indicate that a system with multiple TCP-like flows is globally stable as long as the bandwidth-delay product per flow is not very small.