Using back-pressure to improve TCP performance with many flows

Congestion control of Internet best effort traffic relies mostly on TCP window flow control of individual sessions. This paper argues that such approach does not scale well to a very large number of simultaneously active flows, typical of backbones characterized by large delay-bandwidth products. In this scenario, the TCP window sizes tend to be small and dropping or marking packets alone is not effective to reduce the offered traffic. The analysis presented describes the aggressive TCP dynamics under many flows and suggests that the performance can be improved by applying back-pressure flow control to the aggregate traffic in the backbone. To demonstrated this argument, a link-layer, rate-based, back-pressure mechanism for IP-over-ATM backbones using the ABR service is described. A simulation study of a network with this capability demonstrates the improvement of TCP performance under many flows. The study also considers RED and ECN routers to indicate that these techniques alone are not well positioned to address the many flows scenario either. However, the combination of random early detection (RED) and explicit congestion notification (ECN) routers with back-pressure has the potential to further improve TCP performance.