Generic congestion control through network coordination

With the rapid emergence of new services and the quick change of traffic in Internet, the widely used end-to-end congestion control scheme becomes insufficient. This paper proposes a new congestion control architecture based on the coordination among the core router the edge router, and the host. The core router notifies its immediately upstream nodes when it detects incipient congestion. Those nodes then constrain their sending rates to prevent congestion. No per-flow state information is maintained in routers. The application itself need not take care of the rate adaptation. The architecture provides a uniform solution accommodating both responsive and unresponsive traffic with trivial overhead. It reduces the packet loss to zero, and leads to higher network resource utilization. Response time to congestion is greatly reduced. The architecture has been implemented in the Linux kernel, Some preliminary experiments show its advantages.