A deadlock-free subnetting mechanism for high performance broadcasting in NoC

Abstract

In this letter we present a mechanism to generate subnets for the tasks running on network-on-chip, gaining better performance in broadcast. The present results show that subnetting supplies apparent improvement in the performance of on-chip communications mixed with both unicast and broadcast on manycore processors. However, due to unpredictable task submitting sequence, the applications coming following will be mapped into subnets with different irregular topologies, in which the performance of communications will be restricted by topologies apparently. We propose a flexible subnetting mechanism to build subnet for specified cores on manycore processor to address this problem. Improved hardware support also requires fewer circuits in router design. In this letter, both the unicast and the broadcast traffics are considered. Experiments show that about 10% improvement in performance over the traditional subnetting method is gained on average.