Agent-Based Traffic Merging in Network-on-Chip

Traffic patterns, varying from applications, have great influence on the performance of on-chip network. We observed there exists some balanced traffic, almost distributing traffic load evenly around the whole network. Such traffic pattern makes optimization highly intractable. General optimization method, including modifying the structure of routers and topology, disappointedly, can't give satisfactory results. Moreover, other alternative methods bring good gains, but expose extraordinary high cost. To address this problem, we propose an agent-based traffic merging mechanism in this paper, which reduces the traffic pattern distinctions between applications by creating artificial hotspots at agents and setting express channels between them. With this scheme, part of the communication is shifted to express channels and accelerated to the destination. On the other hand, the pressure of basic network is alleviated to achieve low congestion rate. Furthermore, this mechanism takes fairly low cost of resources and power. Experiments are taken on real applications and a cycle accurate NoC simulator. The results show that when using 8 agents, our method can reduce 23.5% of the latency compared to original 8*8 mesh NoC with nearly no chip area overhead, or reduce 14.1% of the chip area and at the same time reduce 7.9% of the network latency.