Reliability Analysis and Optimization of Embedded Systems using Stochastic Logic and Importance Measures

With shrinking cell geometries, semiconductor devices encounter a wide variety of technical challenges. Among them, the higher vulnerability of these devices to, e. g., neutron-induced soft errors, negative bias temperature instability, short-channel effect, gate leakage, etc. is evident. Reliability is then considered as one of the key requirements in microprocessor systems. In the case of system design, it is thus very important to both evaluate and optimize system reliability together with several other design objectives in the design phase. To target this problem, this work proposes a system-level design methodology, tailored to the design of reliable embedded systems. The proposed methodology models system reliability automatically for both transient and permanent faults. The reliability model of a system implementation is represented by a Success Tree (ST). The ST of a system is extended later in this work to a Success Graph (SG) to support cyclic failure dependencies among the components of a system. The generated ST (or SG) is then evaluated efficiently using stochastic logic. During Design Space Exploration (DSE), thousands of system implementations are explored and evaluated to find the high-quality (Pareto-optimal) implementations. Each of these implementations represents a trade-off given the multiple and often even conflicting design objectives. Therefore, not only each system implementation has to be evaluated automatically, but also has the optimization process to be efficient enough to deal with a huge design space. At this juncture, we propose to augment the optimization process by using the notion of component importance. The idea is to rank the components of a system depending on how they influence the system quality (e. g., reliability). Later for design, enhancement, or maintenance, more resources are dedicated to more important components. Respectively, two novel approaches are proposed in this work to quantify the reliability importance of components following an efficient implementation of them. The main aftermath of these techniques is that they generalize the concept of component importance for both coherent and noncoherent systems. Finally, two concrete optimization approaches are proposed in this work which are guided by the importance of components. The first approach makes local changes to each explored system implementation in order to upgrade (or downgrade) the reliability of each component. The second approach introduces an innovative mutation operator merged into a meta-heuristic opti-