Abstraction techniques for scalable soft error analysis and mitigation

Overview & Synopsis

The main objective of this thesis is to develop techniques that can beused to analyze and mitigate the effects of radiation-induced soft errors in industrialscale integrated circuits. To achieve this goal, several methods have been developedbased on analyzing the design at higher levels of abstraction. These techniquesaddress both sequential and combinatorial SER.Fault-injection simulations remain the primary method for analyzing the effectsof soft errors. In this thesis, techniques which significantly speed-up fault-injectionsimulations are presented. Soft errors in flip-flops are typically mitigated by selectivelyreplacing the most critical flip-flops with hardened implementations. Selectingan optimal set to harden is a compute intensive problem and the second contributionconsists of a clustering technique which significantly reduces the number offault-injections required to perform selective mitigation.In terrestrial applications, the effect of soft errors in combinatorial logic hasbeen fairly small. It is known that this effect is growing, yet there exist few techniqueswhich can quickly estimate the extent of combinatorial SER for an entireintegrated circuit. The third contribution of this thesis is a hierarchical approachto combinatorial soft error analysis.Systems-on-chip are often developed by re-using design-blocks that come frommultiple sources. In this context, there is a need to develop and exchange reliabilitymodels. The final contribution of this thesis consists of an application specificmodeling language called RIIF (Reliability Information Interchange Format). Thislanguage is able to model how faults at the gate-level propagate up to the block andchip-level. Work is underway to standardize the RIIF modeling language as well asto extend it beyond modeling of radiation-induced failures.In addition to the main axis of research, some tangential topics were studied incollaboration with other teams. One of these consisted in the development of a novelapproach for protecting ternary content addressable memories (TCAMs), a specialtype of memory important in networking applications. The second supplementalproject resulted in an algorithm for quickly generating approximate redundant logicwhich can protect combinatorial networks against permanent faults. Finally anapproach for reducing the detection time for errors in the configuration RAM forField-Programmable Gate-Arrays (FPGAs) was outlined.