Failure analysis of distributed scientific workflows executing in the cloud

This work presents models characterizing failures observed during the execution of large scientific applications on Amazon EC2. Scientific workflows are used as the underlying abstraction for application representations. As scientific workflows scale to hundreds of thousands of distinct tasks, failures due to software and hardware faults become increasingly common. We study job failure models for data collected from 4 scientific applications, by our Stampede framework. In particular, we show that a Naive Bayes classifier can accurately predict the failure probability of jobs. The models allow us to predict job failures for a given execution resource and then use these failure predictions for two higher-level goals: (1) to suggest a better job assignment, and (2) to provide quantitative feedback to the workflow component developer about the robustness of their application codes.