Data Mining Based Root-Cause Analysis of Performance Bottleneck for Big Data Workload

Straggler task is commonly considered as the major bottleneck in parallel data processing. Previous work mainly focuses on the coarse-grained straggler detection and optimization such as speculative scheduling. However, fine-grained root-cause analysis of straggler tasks is rarely considered. In addition, existing work simply depends on empirical analysis, which lacks of useful guidance to performance optimization. In this paper, we propose a new methodology of fine-grained straggler root-cause analysis using machine learning. We collect raw metrics from Spark event log and hardware sampling tool, and refine them into high-level metrics for model learning. Then we present the root-cause analysis of stragglers through CART tree. A customized prune method is also applied to improve analysis accuracy. From the analysis, we derive several new findings beyond the well known causes of stragglers. Our work provides a new perspective on identifying and understanding the inefficiency in parallel data processing programs by applying machine learning techniques to fine-grained root-cause analysis of straggler tasks.