An Algorithm-Based Recovery Scheme for Extreme Scale Computing
Hui Liu · 2011
We present an algorithm-based recovery scheme for Exascale computing, which uses both data dependencies and communication-induced redundancies of parallel codes to tolerate fault with low overhead. For some applications, our scheme significantly reduces checkpoint size and introduces no overhead when there is no actual failure in the computation. Fault tolerance Newton's method by tailoring our scheme to the algorithm is performed. Numerical simulations indicate that our scheme introduces much less overhead than diskless check pointing does.