Date of Award

Spring 2012

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational Analysis and Modeling

First Advisor

Chokchai Leangsuksun

Abstract

The failure rate in high-performance computing (HPC) systems continues to escalate as the number of components in these systems increases. This affects the scalability and the performance of parallel applications in large-scale HPC systems. Fault tolerance (FT) mechanisms help mitigating the impact of failures on parallel applications. However, utilizing such mechanisms requires additional overhead. Besides, the overuse of FT mechanisms results in unnecessarily large overhead in the parallel applications. Knowing when and where failures will occur can greatly reduce the excessive overhead. As such, failure prediction is critical in order to effectively utilize FT mechanisms. In addition, it also helps in system administration and management, as the predicted failure can be handled beforehand with limited impact to the running systems.

This dissertation proposes new proficiency metrics for failure prediction based on failure impact in UPC environment that the existing proficiency metrics tire unable to reflect. Furthermore, an efficient log message clustering algorithm is proposed for system event log data preprocessing and analysis. Then, two novel association rule mining approaches are introduced and employed for HPC failure prediction. Finally, the performances of the existing and the proposed association rule mining methods are compared and analyzed.

Share

COinS