Optimized Speculative Execution Strategy for Different Workload Levels in Heterogeneous Spark Cluster

Abstract

Spark is a big data processing framework based on MapReduce, whose calculation model requires that all tasks in all parent stages are completed before starting a new stage. Machine service variability or congested network connections caused by partial or intermittent machine failures become a bottleneck for the Spark framework to execute tasks. In this paper, we focus on the design of speculative execution schemes for heterogeneous Spark from an optimization perspective on different loading conditions. First, we derive the load arrival rate threshold for different operating regimes. Second, for the lightly loaded case, we analyze and propose the speculative execution based on task-cloning algorithm (SETC) which reduce the application completion time by maximizing the overall system utility. Then, for the heavily loaded case, we propose the speculative execution based on straggler-detection algorithm(SESD), which aims to mitigate stragglers. Finally, we conduct experiments to verify the performance of SETC and SESD. Results show that our method is faster than Spark-Speculation, LATE, and SCA by16.7%, 8.2%, and 11.7%. Also it outperforms the baseline algorithms in some metric aspect such as the cluster throughput.