Argonne National Laboratory

Failure Prediction for HPC Systems and Applications: Current Situation and Open Issues

TitleFailure Prediction for HPC Systems and Applications: Current Situation and Open Issues
Publication TypeJournal Article
Year of Publication2013
AuthorsGainaru, A, Cappello, F, Snir, M, Kramer, W
JournalInternational Journal of High Performance Computing Applications
Date Published07/2013
Other NumbersANL/MCS-P5001-0813

As large-scale systems evolve towards post-petascale computing, it is crucial to focus on providing fault-tolerance strategies that aim to minimize fault’s effects on applications. By far the most popular technique is the checkpoint–restart strategy. A complement to this classical approach is failure avoidance, by which the occurrence of a fault is predicted and proactive measures are taken. This requires a reliable prediction system to anticipate failures and their locations. One way of offering prediction is by the analysis of system logs generated during production by large-scale systems. Current research in this field presents a number of limitations that make them unusable for running on real production high-performance computing (HPC) systems. Based on our observations that different failures have different distributions and behaviors, we propose a novel hybrid approach that combines signal analysis with data mining in order to overcome current limitations. We show that by analyzing each event according to its specific behavior, our prediction provides a precision of over 90% and its able to discover about 50% of all failures in a system, result which allows its integration in proactive fault tolerance protocols.