Fine-grained containment domains for throughput processors

Continued scaling of semiconductor technology has made modern processors rely on large design margins to guarantee correct operation under worst case conditions. Design margins appear in the form of higher supply voltage or lower clock frequency, leading to inefficiency. In practice, it is rare to observe such worst-case conditions and the processor can run at a reduced voltage or higher frequency experiencing only few infrequent errors. Recent proposals have used hardware error detectors and recovery mechanisms to detect and re- cover from these rare errors, a technique known as timing speculation. While this is effective for out-of-order processors with inherent capability to recover from misspeculation, implementing similar hardware for throughput processors such as the Graphics Processing Units (GPUs) is prohibitively costly due to the massive amount of thread context that needs to be preserved. Further- more, recovery overhead is much higher since the SIMD (Single Instruction Multiple Data) execution model of GPUs require multiple threads to roll back together in case of an error. In this dissertation, I develop a hardware/software co-design approach to enable reduced-margin operation on GPUs that overcomes the limitations of existing techniques. The proposed scheme leverages the hierarchical programming model of GPUs to provide hierarchical and uncoordinated local checkpoint-recovery. By decomposing a program into a hierarchically nested tree of code blocks which I refer to as containment domains (CDs), the pro- gram becomes amenable to automatic analysis and tuning, and an optimum trade-off can be made between preservation and recovery overhead. To aid this optimization process, an analytical model is developed to estimate the performance efficiency of a given application setting at a given error rate. With the analytical model, an exhaustive search can be performed to find the optimal solution. The tunability also allows the proposed scheme to easily adapt to a wide range of error rates making it future proof against emerging uncertainties in semiconductor design. The proposed scheme combines software and hardware components to achieve the highest efficiency in preservation, restoration, and recovery. The software components include: 1) an API and runtime that lets the programmers describe the hierarchy of containment domains within an application and preserve the state required for rollback recovery, and 2) a compiler analysis that automatically inserts preservation routines for register variables. The hardware components include: 1) a stack structure to keep track of recovery program counters (PC), 2) a set of error containment mechanisms to guarantee that no erroneous data is propagated outside of a containment domain and 3) an error reporting architecture that keeps track of affected threads and initiate recovery of them.