Odd-even based adaptive two-way routing in mesh NoCs for hotspot mitigation

Abstract

Network-on-Chip is adapted as a profitable framework for communication in on-chip multiprocessors. Congestion management using adaptive routing techniques become the major research focus in recent days. Hotspots are congested cores in multi-core systems, which has to deal with large amount of packetized data than other cores in the network. When a packet has to pass through hotspots, it will adversely affect the overall system performance. We identify the hotspot cores using counters, and propose an adaptive two-way routing algorithm by restricting some routes as in odd-even turn model to handle the presence of hotspots. The algorithm designed not only de-routes packets from current hotspots, but also reduces the possibility of nearby hotspots formation in the future. Experimental results using SPEC 2006 CPU benchmarks show that the chances of hotspots are large in highly congested traffic, and our algorithm gives about 14% average reduction in packet latency than existing routing methods in presence of hotspot cores.