Towards machine learning-based self-tuning of Hadoop-Spark system

Abstract

Apache Spark is an open source distributed platform which uses the concept of distributed memory for processing big data. Spark has more than 180 predominant configuration parameter. Configuration settings directly control the efficiency of Apache spark while processing big data, to get the best outcome yet a challenging task as it has many configuration parameters.  Currently, these predominant parameters are tuned manually by trial and error. To overcome this manual tuning problem in this paper proposed and developed a self-tuning approach using machine learning. This approach can tune the parameter value when it’s required. The approach was implemented on Dell server and experiment was done on five different sizes of the dataset and parameter. A comparison is provided to highlight the experimented result of the proposed approach with default Spark configuration system. The results demonstrate that the execution is speeded-up by about 33% (on an average) compared to the default configuration.