Hadoop framework implementation and performance analysis on a cloud

Hadoop framework uses MapReduce programming paradigm to process big data by distributing data across a cluster and aggregating. MapReduce is one of the methods used to process big data hosted on large clusters. In this method, jobs are processed by dividing into small pieces and distributing over nodes. Parameters such as distributing method over nodes, the number of jobs held in a parallel fashion and the number of nodes in the cluster affect the execution time of jobs. The aim of this paper is to determine how number of nodes, maps and reduces affect the performance of Hadoop framework on a cloud environment. For this purpose, tests were carried out on a Hadoop cluster with 10 nodes hosted on a cloud environment by running PiEstimator, Grep, Teragen and Terasort benchmarking tools on it. These benchmarking tools available under Hadoop framework are classified as CPU-intensive and CPU-light applications as a result of tests. In CPU-light applications; increasing number of nodes, maps and reduces do not improve efficiency of these applications, even they cause increase of time spent on jobs by using system resources unnecessarily. Therefore, in CPU-light applications, selecting number of nodes, maps and reduces as minimum are found as the optimization of time spent on a process. In CPU-intensive applications, according to the phase that small job pieces are processed, it is found that selecting number of maps or reduces equal to total number of CPUs on a cluster as the optimization of time spent on a process.-

Keywords: Big Data, Hadoop, Cloud Computing, MapReduce, KVM Virtualization Environment, Benchmarking Tools, Ganglia.