Hessenberg reduction with transient error resilience on GPU-based hybrid architectures

Graphics Processing Units (GPUs) have been seeing widespread adoption in thefield of scientific computing, owing to the performance gains provided oncomputation-intensive applications. In this paper, we present the design andimplementation of a Hessenberg reduction algorithm immune to simultaneoussoft-errors, capable of taking advantage of hybrid GPU-CPU platforms. Thesesoft-errors are detected and corrected on the fly, preventing the propagationof the error to the rest of the data. Our design is at the intersection betweenseveral fault tolerant techniques and employs the algorithm-based faulttolerance technique, diskless checkpointing, and reverse computation to achieveits goal. By utilizing the idle time of the CPUs, and by overlapping bothhost-side and GPU-side workloads, we minimize the resilience overhead. Experimental results have validated our design decisions as our algorithmintroduced less than 2% performance overhead compared to the optimized, butfault-prone, hybrid Hessenberg reduction.