Cheetah: A Dynamic Performance Optimization Approach on Heterogeneous Big Data Analytics Cluster

Any MapReduce-based big data analytics clusters systems like Hadoop, Spark face the substantial challenge of the Long Tail problem: a small subset of straggling tasks significantly impede parallel jobs completion. Specially, in heterogeneous environments some tasks become stragglers because of poor performance of some computing nodes, data skew, etc. Therefore, stragglers are well recognized as a major bottleneck in big data processing and hence the early detection and accurate identification of stragglers can have important impacts on big data processing. In past years speculative execution strategies that have been proposed to meet their challenges such as misjudgment or delaying of straggling tasks, improper selection of backup nodes, etc., which result in inaccurate and inefficient performance of the speculative execution. Encouraged by recent successes in applying reinforcement learning (RL) techniques to solve complex online control problems, we study RL can be used for automatic choose right strategy to address stragglers without human-intervention. In this paper, we present Cheetah: a novel dynamic optimal approach for speculative strategy which identifies stragglers by reinforcement learning and automatic chooses the best strategy to launch speculative tasks on heterogeneous cluster. We implement Cheetah with popular reinforcement learning frameworks, and deploy it on a testbed of heterogeneous Spark cluster with 4 nodes. According to the experimental results, Compared to existing approaches, Cheetah reduces the job completion time in different type applications while achieving superior performance. For example, it demonstrates up to almost 16% reduction average job completion time and improve 25% on accuracy over existing solutions on the heterogeneous Spark cluster.