Congestion Control Based on Transmission Times

Abstract

Congestion leads to a severe performance degradation in multiprocessor interconnection networks. Therefore, the use of techniques that prevent network saturation are of crucial importance to avoid high execution times. In this paper, we propose a new mechanism that uses only local information to avoid network saturation in wormhole networks. In order to detect congestion, each network node computes the quotient between the real transmission time of messages and its minimum theoretical value. If this ratio is greater than a threshold, the physical channel used by the message is considered congested. Depending on the number of congested channels, the available bandwidth to inject messages is reduced. The main contributions of the new mechanism are three: i) it can detect congestion in a remote way, but without transmitting control information through the network; ii) it tries to dynamically adjust the effective injection bandwidth available at each node; and iii) it is starvation-free. Evaluation results show that the proposed mechanism avoids network performance degradation for different network loads and topologies. Indeed, the mechanism does not introduce any penalty for low and medium network loads, where no congestion control mechanism is required.