Handling global traffic in future CMP NoCs

It was recently shown that if computation locality is properly employed, Chip-Multi-Processors (CMP) traffic patterns can be modeled with a bandwidth version of Rent´s rule. The Communication Probability Distributions (CPD) derived from the Rent´s rule imply that most end to end packets are exchanged by nearest-neighbors. We show that while packets exchanged with nearest-neighbor dominate peer to peer traffic, their contribution to the overall NoC traffic decreases rapidly as the system grows. Correspondingly, the bandwidth consumed by long distance packets (a. ka. global packets) becomes dominant starting from medium-size systems, despite their low injection rate. To accommodate this phenomenon, we introduce PyraMesh - a novel family of multilevel hierarchical 2D mesh topologies resembling a pyramid structure. In PyraMesh, global packets are separated from the local ones and routed through the upper levels of the hierarchy. PyraMesh is shown to improve light-load latencies and raise the saturation point of the network to higher injection rates, as compared with previously presented NoC topologies.