Algorithm-Based Recovery for Newton´s Method without Checkpointing

Check pointing is the most popular fault tolerance method used in high-performance computing (HPC) systems. However, increasing failure rates requires more frequent checkpoints, thus makes check pointing more expensive. We present a checkpoint-free fault tolerance technique. It takes advantage of both data dependencies and communication-induced redundancies of parallel applications to tolerate fail-stop failures. Under the specified conditions, our technique introduces no additional overhead when there is no actual failure in the computation and recover the lost data with low overhead. We add fault-tolerant capacity to Newton´s method by using our scheme and diskless check pointing. Numerical simulations indicate that our scheme introduces much less overhead than diskless check pointing does.