An efficient approach to optimise I/O cost in data-intensive applications using inverted indexes on HDFS splits

Abstract

Hadoop is prominent for its distributed file system (HDFS) and scalability. Hadoop MapReduce framework is extensively used in big data analytics and business-intelligence applications. The analytic queries executed by these applications often include multiple ad hoc queries and aggregate queries with some selection predicates. The cost of executing these queries grows incredibly as the size of dataset grows. The most effective strategy to improve query performance in such applications is to process only relevant data keeping irrelevant data aside, which can be done using index structures. This paper is an attempt to improve query performance by avoiding full scans on data files. The algorithms used in this paper create inverted indexes on HDFS input splits. We show how query processing in MR jobs can benefit in terms of performance by employing these custom inverted indexes. The experiments demonstrate that queries executed using indexed data execute 1.5x faster than the traditional queries.