Adaptive and Power-Aware Resilience for Extreme-scale Computing

Abstract

With concerted efforts from researchers in hardware, software, algorithm,, resource management, HPC is moving towards extreme-scale, featuring a computing capability of exaFLOPS. As we approach the new era of computing, however, several daunting scalability challenges remain to be conquered. Delivering extreme-scale performance will require a computing platform that supports billion-way parallelism, necessitating a dramatic increase in the number of computing, storage,, networking components. At such a large scale, failure would become a norm rather than an exception, driving the system to significantly lower efficiency with unprecedented amount of power consumption. To tackle these challenges, we propose an adaptive, power-aware algorithm, referred to as Lazy Shadowing, as an efficient, scalable approach to achieve high-levels of resilience, through forward progress, in extreme-scale, failure-prone computing environments. Lazy Shadowing associates with each process a (process) that executes at a reduced rate,, opportunistically rolls forward each shadow to catch up with its leading process during failure recovery. Compared to existing fault tolerance methods, our approach can achieve 20% energy saving with potential reduction in solution time at scale.