| With continued CMOS scaling, future shipped hardware will be increasingly vulnerable to faults and inevitably fail in the field for a variety of reasons such as aging or wear-out, radiation, infant mortality due to inadequate burn-in, design defects, manufacturing defects, and so on. Further, this reliability threat is expected to pervade even the mainstream computing market, making traditional solutions that involve redundancy in space, time, and/or information too expensive to be broadly deployable. Hence, there is a need for effective in-field reliability solutions that incur low overheads in area, power, and performance and handle multiple sources of errors.;The main contribution of this dissertation is to investigate the design of a low-cost comprehensive reliability solution that detects, diagnoses, and recovers from in-field errors. Our design is based on the following two key observations. (1) Hardware reliability solutions only need to handle device faults that manifest to higher levels of the system and appear as software bugs. (2) Despite the growing reliability problem, the fault-free operation remains the common case and must be optimized.;These insights drive the design of a novel reliability solution that employs near zero overhead "always-on" monitors to detect hardware faults by watching for the anomalous software behavior (called symptoms). After a detection, a potentially expensive diagnosis algorithm is invoked to diagnose the source of the error and ensure full recovery. While the diagnosis may incur high overhead, it is only invoked in the rare case of a detection. We believe that the very low cost detection coupled with higher cost diagnosis is the right tradeoff for achieving very low cost in reliability solutions.;With these strategies, this dissertation presents a comprehensive reliability solution, called SWAT (Soft Ware Anomaly Treatment), that detects, diagnoses, and recovers from in-the-field failures at very low cost. For hardware error detection, SWAT relies on low cost, "always-on" monitors of software symptoms. After a detection, SWAT uses a novel technique called trace based fault diagnosis to identify whether the symptom detection is a result of a hardware or software error and to diagnose the faulty microarchitectural component in case of a hardware permanent fault. For recovery, SWAT aims to use hardware checkpointing for restoring the fault-free execution state and relies on output event buffering for preventing hardware faults from propagating and becoming visible outside of the system.;We evaluated the SWAT system with statistical fault injection experiments. Our results show that simple monitors of software symptoms can achieve high permanent hardware fault detection coverage and comparable transient fault detection coverage as previously proposed symptom-based transient fault detection schemes. After a detection, our traced based nucroarchitecture-level diagnosis correctly identifies most of the detected hardware permanent faults, facilitating fine-grained repair. For error recovery, we found that high system recoverability can only be attained by employing both hardware checkpointing and output buffering mechanisms, indicating that both are essential for ensuring hardware reliability. These results show empirically that the SWAT system is an effective hardware reliability solution that detects, diagnoses, and recovers from failed components in the field.;The final contribution of this dissertation is to investigate the accuracy of microarchitecture-level fault modeling. To achieve this goal, we present a novel fault simulation framework called SWAT-Sim that can model gate-level faults accurately while achieving speed comparable to microarchitectural simulations. Using SWAT-Sim, we found that existing microarchitecture-level faults cannot, in general, accurately represent gate-level faults. The SWAT-Sim framework, therefore, serves as an important research vehicle for both SWAT and other ongoing or future research in hardware reliability.;In summary, this dissertation shows, for the first time, that a comprehensive hardware reliability solution can be realized by treating the software-level symptoms caused by both permanent and transient hardware faults. The presented work lays the foundation for the SWAT approach and paves the way for future work on low-cost software anomaly based resilient systems. |
