DISSERTATION Fault-Tolerant Distributed Algorithms for On-Chip Tick Generation: Concepts, Implementations and Evaluations

In the course of this thesis a novel approach for the on-chip generation of a fault-tolerant clock is developed. At first this is motivated by the fact that with shrinking feature sizes and the accompanying increase of transient failure rates it is more and more desirable to provide VLSI (Very Large Scale Integration) circuits that incorporate mechanisms for fault tolerance. In particular, the conducted research concentrates on the most prominent single point of failure of modern chip design, namely, the clock signal of synchronous circuits. After surveying alternative design approaches and existing schemes for achieving fault tolerance a novel fault-tolerant clocking scheme is introduced. The proposed clock generation method is based on the hardware implementation of a well known distributed clock synchronization algorithm. Most notably, it provides scalable fault tolerance for up to f arbitrary (Byzantine) failures in a system of n ≥ 3f + 2 tick generation nodes. Additionally, the clocking scheme’s operation does not rely on the synchronization of clock sources, like quartz oscillators; in fact, the distributed clock signals are generated in a synchronized way. This unique property relieves the design from metastability issues at clock boundaries. The transformation of the original software-based algorithm to the peculiarities of chip design proved to be an intricate task. Therefore, the major part of the work deals with the design and development process of the algorithm’s hardware equivalent finally resulting in a fully operational VLSI chip design. To assess the properties of the novel fault-tolerant clocking approach and to show its feasibility exhaustive evaluations have been performed. The presented assessments aim at a thorough characterization of (i) the developed chip design and (ii) the distributed clock generation scheme on which these chips are based. Additionally, the conducted measurements allowed to validate worst-case measures which were derived in advance from the formal analysis of the clocking approach. In order to attain a more comprehensive characterization of the design, the presented worst-case evaluations have been supported by measurements and simulations for typical operating scenarios. The presented work concludes with a short summary and a brief treatment of the most notable topics for ongoing and future research.