Abstract
Research problem: The impact of failures on software systems can be substantial since the recovery process can require unexpected amounts of time and resources. Accurate failure predictions can help in mitigating the impact of failures. Resources, applications, and services can be scheduled to limit the impact of failures. However, providing accurate predictions sufficiently ahead is challenging. Log files contain messages that represent a change of system state. A sequence or a pattern of messages may be used to predict failures. Contribution: We describe an approach to predict failures based on log files using Random Indexing (RI) and Support Vector Machines (SVMs).
Method: RI is applied to represent sequences: each operation is characterized in terms of its context. SVMs associate sequences to a class of failures or non-failures. Weighted SVMs are applied to deal with imbalanced datasets and to improve the true positive rate. We apply our approach to log files collected during approximately three months of work in a large European manufacturing company.
Results: According to our results, weighted SVMs sacrifice some specificity to improve sensitivity. Speci- ficity remains higher than 0.80 in four out of six analyzed applications.
Conclusions: Overall, our approach is very reliable in predicting both failures and non-failures.