On Fault Resilient Network-on-Chip for Many Core Systems

Rapid scaling of transistor gate sizes has increased the density of on-chip integration and paved the way for heterogeneous many-core systems-on-chip, significantly improving the speed of on-chip processing. The design of the interconnection network of these complex systems is a challenging one and the network-on-chip (NoC) is now the accepted scalable and bandwidth efficient interconnect for multi-processor systems on-chip (MPSoCs). However, the performance enhancements of technology scaling come at the cost of reliability as on-chip components particularly the network-on-chip become increasingly prone to faults. In this thesis, we focus on approaches to deal with the errors caused by such faults. The results of these approaches are obtained not only via time-consuming cycle-accurate simulations but also by analytical approaches, allowing for faster and accurate evaluations, especially for larger networks. Redundancy is the general approach to deal with faults, the mode of which varies according to the type of fault. For the NoC, there exists a classification of faults into transient, intermittent and permanent faults. Transient faults appear randomly for a few cycles and may be caused by the radiation of particles. Intermittent faults are similar to transient faults, however, differing in the fact that they occur repeatedly at the same location, eventually leading to a permanent fault. Permanent faults by definition are caused by wires and transistors being permanently short or open. Generally, spatial redundancy or the use of redundant components is used for dealing with permanent faults. Temporal redundancy deals with failures by re-execution or by retransmission of data while information redundancy adds redundant information to the data packets allowing for error detection and correction. Temporal and information redundancy methods are useful when dealing with transient and intermittent faults. In this dissertation, we begin with permanent faults in NoC in the form of faulty links and routers. Our approach for spatial redundancy adds redundant links in the diagonal direction to the standard rectangular mesh topology resulting in the hexagonal and octagonal NoCs. In addition to redundant links, adaptive routing must be used to bypass faulty components. We develop novel fault-tolerant deadlock-free adaptive routing algorithms for these topologies based on the turn model without the use of virtual channels. Our results show that the hexagonal and octagonal NoCs can tolerate all 2-router and 3-router faults, respectively, while the mesh has been shown to tolerate all 1-router faults. To simplify the restricted-turn selection process for achieving deadlock freedom, we devised an approach based on the channel dependency matrix instead of the state-of-the-art Duato's method of observing the channel dependency graph for cycles. The approach is general and can be used for the turn selection process for any regular topology. We further use algebraic manipulations of the channel dependency matrix to analytically assess the fault resilience of the adaptive routing algorithms when affected by permanent faults. We present and validate this method for the 2D mesh and hexagonal NoC topologies achieving very high accuracy with a maximum error of 1%. The approach is very general and allows for faster evaluations as compared to the generally used cycle-accurate simulations. In comparison, existing works usually assume a limited number of faults to be able to analytically assess the network reliability. We apply the approach to evaluate the fault resilience of larger NoCs demonstrating the usefulness of the approach especially compared to cycle-accurate simulations. Finally, we concentrate on temporal and information redundancy techniques to deal with transient and intermittent faults in the router resulting in the dropping and hence loss of packets. Temporal redundancy is applied in the form of ARQ and retransmission of lost packets. Information redundancy is applied by the generation and transmission of redundant linear combinations of packets known as random linear network coding. We develop an analytic model for flexible evaluation of these approaches to determine the network performance parameters such as residual error rates and increased network load. The analytic model allows to evaluate larger NoCs and different topologies and to investigate the advantage of network coding compared to uncoded transmissions. We further extend the work with a small insight to the problem of secure communication over the NoC. Assuming large heterogeneous MPSoCs with components from third parties, the communication is subject to active attacks in the form of packet modification and drops in the NoC routers. Devising approaches to resolve these issues, we again formulate analytic models for their flexible and accurate evaluations, with a maximum estimation error of 7%