Avoiding Network Congestion with Local Information

Congestion leads to a severe performance degradation in multiprocessor interconnection networks. Therefore, the use of techniques that prevent network saturation are of crucial importance. Some recent proposals use global network information, thus requiring that nodes exchange some control information, which consumes a far from negligible bandwidth. As a consequence, the behavior of these techniques in practice is not as good as expected.In this paper, we propose a mechanism that uses only local information to avoid network saturation. Each node estimates traffic locally by using the percentage of free virtual output channels that can be used to forward a message towards its destination. When this number is below a threshold value, network congestion is assumed to exist and message throttling is applied. The main contributions of the proposed mechanism are two: i) it is more selective than previous approaches, as it only prevents the injection of messages when they are destined to congested areas; and ii) it outperforms recent proposals that rely on global information.