Probabilistic accuracy bounds for fault-tolerant computations that discard tasks

摘要: We present a new technique for enabling computations to survive errors and faults while providing bound on any resulting output distortion. A developer using the first partitions computation into tasks. The execution platform then simply discards task that encounters an error or fault completes by executing remaining This can substantially improve robustness of in face faults. potential concern is discarding tasks may change result produces.Our randomly samples executions program at varying failure rates obtain quantitative, probabilistic model characterizes distortion as function rates. By bounds distortion, allows users confidently accept results produced with failures long falls within acceptable bounds. approach prove be especially useful successfully hardware distributed computing environments.Our also produces timing time combination models quantifies accuracy/execution tradeoff. It therefore enables development techniques purposefully fail reduce keeping