MUGEN: A high-performance fault-tolerant routing algorithm for unreliable Networks-on-Chip

Abstract

NoCs (Networks-on-Chip) are an attractive alternative to communication buses for SoCs (Systems-on-Chip) as they offer both high scalability and low power consumption. However, designing such systems in the nanoscale era brings up some serious concerns about reliability. Our aim is to design robust NoCs while limiting performance degradation. In this paper, we introduce several techniques meant to increase the reliability and performance of NoCs. We combine these techniques to build a fault-tolerant, deadlock-free and congestion-aware routing algorithm called MUGEN. The algorithm comprises an optimized method to exchange messages between different virtual channel classes, a selection function that uses distant router link information to avoid dead-ends and a new congestion metric used to guide routing decisions towards less congested areas. We simulate an 8×8 Mesh NoC with fault injection to evaluate each method used by MUGEN individually before comparing the full algorithm with existing works from literature. We present promising results about the proposed techniques both in terms of fault-tolerance and performance.