Abstract
In this paper we develop a general approach to global reactive congestion control in multicomputer networks. The approach uses a timeout mechanism to detect congestion, and exploits control lines such as those used for handshaking in the flit-level flow control of wormhole routers to distribute information about congestion. It is also based on a mechanism that limits the demands placed by the network interface and the processing element. The approach is described in detail and evaluated through simulation experiments. We show that the proposed congestion control can provide network stability and predictable network performance. By choosing the right timeout, we can provide bounds on average delay and worst-case delay. Furthermore, with appropriate timeouts the network can be kept out of saturation. Other attributes of the approach include fairness and applicability to a wide range of network architectures.