An OpenMP Based Approach for Parallelization and Performance Evaluation of k-Means Algorithm

Abstract

In today’s digital world, the volume of data is drastically increasing due to the continuous flow of data from various heterogenous sources such as WWW, social media, environmental sensors, huge enterprise data warehouses, bioinformatic labs etc. to name a few. This results in creation of many high-volume datasets in various domains. Processing such large datasets is a tedious task, therefore they need to be categorized into smaller subsets using various supervised or unsupervised classification techniques. Clustering is the process of statistically analyzing and categorizing data objects with similarity, into substantially homogeneous groups, called data clusters. k-Means is the most common, simple and popular clustering technique, due to its ease of implementation, usability and wide range of applications. One of the issues associated with the k-Means algorithm is that it suffers from the scalability problem due to which, its performance degrades as the dataset sizes grow.  In order to address this issue, we have presented an OpenMP based parallelized k-means algorithm which results in better computational cost as compared with its sequential counterpart. Computational performance results of both sequential and OpenMP based k-means algorithms are illustrated and compared.