On resilient task allocation and scheduling with uncertain quality checkers

摘要

Many emerging applications are inherently error-resilient and hence do not require exact computation. Previous work on resilience-aware task allocation and scheduling problem first generates an initial energy-efficient task schedule on voltage-scalable multiprocessor system at design-time, and then conducts voltage adjustment according to runtime quality checking result. While energy efficiency improvements are quite encouraging, the final quality requirement might be violated because quality checkers are usually designed based on partial information and they are not always correct. In this paper, we propose to address the uncertainty issue of quality checkers in resilient task allocation and scheduling. To be specific, given the initial task schedule and quality checkers, we propose (i) a solution that ensures the final output quality with maximized probability, which can be applied to any approximate computing quality management systems; and (ii) a greedy runtime algorithm to achieve optimized energy efficiency gains. Experimental results on various task graphs demonstrate the efficacy of our proposed technique.