DescriptionThe growing computational and storage needs of scientific applications mandate the deployment of extreme-scale parallel machines, such as IBM's BlueGene/L, a 64K dual-core processor system. One of the challenges of designing and deploying such systems in a production setting is the need to take failure occurrences into account. Once the large scale system equipped with a failure predictability, the fault tolerance and resource management strategies of the system can be improved significantly, and its performance can be highly increased.
This dissertation is based on the Reliability, Availability and Serviceabilit (RAS) events generated by IBM BlueGene/L over a period of 142 days. Using these logs, we performed failure analysis, modeling, and prediction. Filters are created to reveal the system failure behaviors, three preliminary models are identified for the failures, and finally, three failure predictors are established for the system. We heavily use data mining and time series analysis techniques for this dissertation. Our comprehensive evaluation demonstrates that our Bi-Modal Nearest Neighbor predictor greatly outperforms the other two (RIPPER and LIBSVM based), leading
to an F-measure of 70% and 50% for a 12-hour and 6-hour prediction window size.