A Fine-Grained Source-Throttling Method for Mesh Architectures

Source-throttling algorithms aim at adjusting network load appropriately so that high network throughput can be maintained and latency can be decreased efficiently. One challenge of designing a source-throttling algorithm is how to precisely evaluate network congestion status. State-of-the-art source-throttling algorithms for mesh-connected network-on-chip usually evaluate congestion status of the whole network or only that of neighbor routers. They usually appear empirical or locally-greedy. To avoid such problems, we propose a fine-grained source-throttling algorithm. Our main contributions are: 1, we can quantify the impact of throttling action on network routers; 2, we can monitor network traffic more intelligently compared with previous work, only the routers that are most affected by the source's throttling action are used to evaluate congestion status of the RSD network. Based on such measures, most throttling parameters shall not be empirically set but be precisely figured out by a smarter searching algorithm of finding anchor routers as the throttling baseline. Thus, it can solve over-throttling and under-throttling problems more precisely. Time complexity of proposed method is only O(M * N) (M*N is the scale of 2D mesh network). Simulation results show that it has higher throughput, lower fluctuation and lower latency than source-throttling algorithms with empirical parameters such as INC and self-tuned technologies under different network scenarios.