Best Effort Session-Level Congestion Control

Congestion caused by a large number of interacting TCP flows at a bottleneck network link is different from that caused by a lesser number of flows sending large amounts of data -the former would require cutting down the number of competing flows, while cutting down the data sending rate is sufficient for the latter. However, since existing congestion control schemes view congestion only from a packet-level perspective, they treat both to be the same, resulting in suboptimal performance. We propose two best effort, search-based, session (or flow) level congestion control strategies for the Internet, to complement existing packet-level congestion control schemes. Our strategies control the number of competing flows to optimize for the flow completion rate and the flow completion time. Furthermore, our session control mechanisms do not require any per-flow state or computation at the routers, make no assumption about input traffic characteristics and requirements, avoid starvation of new flows when existing flows do not leave the system, and do not require any end host TCP modifications. Using evaluations under a wide variety of static and varying traffic load conditions, we demonstrate the significant performance and fairness gains that our session control mechanisms provide.