Book Description

Continuous improvements in transistor scaling together with microarchitectural advances have made possible the widespread adoption of high-performance processors across all market segments. However, the growing reliability threats induced by technology scaling and by the complexity of designs are challenging the production of cheap yet robust systems. Soft error trends are haunting, especially for combinational logic, and parity and ECC codes are therefore becoming insufficient as combinational logic turns into the dominant source of soft errors. Furthermore, experts are warning about the need to also address intermittent and permanent faults during processor runtime, as increasing temperatures and device variations will accelerate inherent aging phenomena. These challenges specially threaten the commodity segments, which impose requirements that existing fault tolerance mechanisms cannot offer. Current techniques based on redundant execution were devised in a time when high penalties were assumed for the sake of high reliability levels. Novel light-weight techniques are therefore needed to enable fault protection in the mass market segments. The complexity of designs is making post-silicon validation extremely expensive. Validation costs exceed design costs, and the number of discovered bugs is growing, both during validation and once products hit the market. Fault localization and diagnosis are the biggest bottlenecks, magnified by huge detection latencies, limited internal observability, and costly server farms to generate test outputs. This thesis explores two directions to address some of the critical challenges introduced by unreliable technologies and by the limitations of current validation approaches. We first explore mechanisms for comprehensively detecting multiple sources of failures in modern processors during their lifetime (including transient, intermittent, permanent and also design bugs). Our solutions embrace a paradigm where fault tolerance is built based on exploiting high-level microarchitectural invariants that are reusable across designs, rather than relying on re-execution or ad-hoc block-level protection. To do so, we decompose the basic functionalities of processors into high-level tasks and propose three novel runtime verification solutions that combined enable global error detection: a computation/register dataflow checker, a memory dataflow checker, and a control flow checker. The techniques use the concept of end-to-end signatures and allow designers to adjust the fault coverage to their needs, by trading-off area, power and performance. Our fault injection studies reveal that our methods provide high coverage levels while causing significantly lower performance, power and area costs than existing techniques. Then, this thesis extends the applicability of the proposed error detection schemes to the validation phases. We present a fault localization and diagnosis solution for the memory dataflow by combining our error detection mechanism, a new low-cost logging mechanism and a diagnosis program. Selected internal activity is continuously traced and kept in a memory-resident log whose capacity can be expanded to suite validation needs. The solution can catch undiscovered bugs, reducing the dependence on simulation farms that compute golden outputs. Upon error detection, the diagnosis algorithm analyzes the log to automatically locate the bug, and also to determine its root cause. Our evaluations show that very high localization coverage and diagnosis accuracy can be obtained at very low performance and area costs. The net result is a simplification of current debugging practices, which are extremely manual, time consuming and cumbersome. Altogether, the integrated solutions proposed in this thesis capacitate the industry to deliver more reliable and correct processors as technology evolves into more complex designs and more vulnerable transistors.