Communication optimisation for intermediate data of MapReduce computing model

Abstract

MapReduce is a typical computing model for processing and analysis of big data. MapReduce computing job produces a large amount of intermediate data after map phase. Massive intermediate data results in a large amount of intermediate data communication across rack switches in the Shuffle process of MapReduce computing model, this degrades the performance of heterogeneous cluster computing. In order to optimise the intermediate data communication performance of map-intensive jobs, the characteristics of pre-running scheduling information of MapReduce computing jobs are extracted, and job classification is realised by machine learning. The jobs of active intermediate data communication are mapped into a rack to keep the communication locality of intermediate data. The jobs with inactive communication are deployed to the nodes sorted by computing performance. The experimental results show that the proposed communication optimisation scheme has a good effect on Shuffle-intensive jobs, and can reach 4%–5%. In the case of larger amount of input data, the communication optimisation scheme is robust and can adapt to heterogeneous cluster. In the case of multi-user application scene, the intermediate data communication can be reduced by 4.1%.