Optimizing data locality by executor allocation in spark computing environment

Abstract

Data locality is an important concept in big data processing. Most of the existing research optimized data locality from the aspect of task scheduling. However, as the execution container of tasks, the executors started on which nodes can directly affect the locality level achieved by the tasks. This paper tries to improve the data locality by executor allocation for reduce stage in Spark computing environment. Firstly, we calculate the network distance matrix of executors and formulate an optimal executor allocation problem to minimize the total communication distance. Then, when the network distance between executors satisfies the triangular inequality, an approximate algorithm is proposed; and when the network distance between executors does not satisfy the triangular inequality, a greedy algorithm is proposed. Finally, we evaluate the performance of our algorithms in a practical Spark cluster by using several representative micro-benchmarks (Sort and Join) and macro-benchmarks (PageRank and LDA). Experimental results show that the proposed algorithms can decrease the execution time of tasks for lower data communication.