Fault Management in Map-Reduce Through Early Detection of Anomalous Nodes

Map-Reduce frameworks such as Hadoop have built-in fault-tolerance mechanisms that allow jobs to run to completion even in the presence of certain faults. However, these jobs can experience severe performance penalties under faulty conditions. In this paper, we present Fault-Managed Map-Reduce (FMR) which augments Hadoop with the functionality to mitigate job execution time penalties. FMR uses an anomaly detection algorithm based on sparse coding to anticipate a faulty slave node. This proposed technique has the following key advantages: (1) model training uses only normal-class data, (2) time taken for prediction is less than a second, and (3) confidence estimates are produced along with the anomaly prediction. FMR uses the result of anomaly detection to invoke a closed-loop recovery action, namely dynamic resource scaling. A scaling heuristic is proposed to determine the extent of scaling necessary to reduce impending performance penalty. FMR facilitates practical adoption by being implemented as a set of libraries and scripts that require no changes to the underlying source code of Hadoop. A set of realistic Map-Reduce applications were studied through a few thousand job executions on a 72-node Hadoop testbed. Detailed empirical evaluation shows that FMR successfully mitigates performance penalties from 119% down to 14%, averaged across experiments.