A Combined Analytical Modeling Machine Learning Approach for Performance Prediction of MapReduce Jobs in Hadoop Clusters

Nowadays MapReduce and its open source implementation, Apache Hadoop, are the most widespread solutions for handling massive dataset on clusters of commodity hardware. At the expense of a somewhat reduced performance in comparison to HPC technologies, the MapReduce framework provides fault tolerance and automatic parallelization without any efforts by developers. Since in many cases Hadoop is adopted to support business critical activities, it is often important to predict with fair confidence the execution time of submitted jobs, for instance when SLAs are established with end-users. In this work, we propose and validate a hybrid approach exploiting both queuing networks and support vector regression, in order to achieve a good accuracy without too many costly experiments on a real setup. The experimental results show how the proposed approach attains a 21% improvement in accuracy over applying machine learning techniques without any support from analytical models.