Towards Dependable Network-on-Chip Architectures

Abstract

The aggressive semiconductor technology scaling provides the means for doubling the amount of transistors on a single chip each and every 18 months. To efficiently utilize these vast chip resources, Multi-Processor Systems on Chip (MPSoCs) integrated with a Network-on-Chip (NoC) communication infrastructure have been widely investigated. However, the transistor miniaturization also significantly increases the possibility of transient and permanent faults occurrence inside the chip, especially for NoCs as they geometrically spread all over the chip real estate. To provide dependable communication service, the NoC must maintain its functionality and gracefully degrade its performance in the presence of faults. In this dissertation, we propose several novel NoC tailored mechanisms to tolerate faults induced by, e.g., variability agents, ageing, environmental aggression factors, as well as to efficiently utilize still functional NoC components. We first introduce a low cost method to allow for correct flit transmission even when soft errors are occurring in the router control plane. Then we propose a Flit Serialization (FS) strategy to tolerate broken link wires and to efficiently utilize the remaining link bandwidth. Within the FS framework heavily defected links whose fault levels exceed a certain threshold value are deactivated to diminish the congestion in their upstream routers. Moreover, we design a distributed logic based routing algorithm able to tolerate totally broken links as well as to efficiently utilize UnPaired Functional (UPF) Links in partially defected interconnects. We also introduce a link bandwidth aware run-time task mapping algorithm to improve the mapping quality for newly injected applications in the MPSoCs. Last but not least, we discuss the application of aforementioned strategies in 3D NoC systems and propose a Bus Virtual channel Allocation (BVA) mechanism to enable vertical wormhole switching to improve the performance of 3D NoC-Bus hybrid systems. All proposals are evaluated in our mixed language NoC simulation platform and their advantage over state of the art counterparts are proved by means of experimental results.