Dynamic migratory selection strategy for adaptive routing in mesh NoCs

Abstract

NoC architectures are the most commonly used communication framework for multicore processors. Few factors that affect the performance of an on-chip network include the efficiency of the routing algorithm and the effectiveness of the output selection strategy used. All popular selection strategies use a static technique that behaves uniformly across various traffic patterns. This paper proposes a cost effective adaptive model of Regional Congestion Awareness (RCA) selection strategy. The traffic analyser incorporated in the proposed model learns the flit-flow pattern at each router and makes RCA behaves like the local best selection strategy under local traffic. Under non-local traffic, the normal RCA selection strategy works as it is. This switching (migration) between two selection strategies is done by proper controlling of aggregation and propagation mechanisms of RCA. As only in-router information is used for this switching, the design has no additional communication overhead. This dynamic switching decreases average packet latency and effectively optimises the network resources depending on traffic pattern, thereby, reducing power consumption. Our experiments on 8°8 mesh NoC with various synthetic and real traffic patterns show promising improvements compared to the existing baseline adaptive selection strategies.