Improving I/O Efficiency in Hadoop-Based Massive Data Analysis Programs

摘要

Apache Hadoop has been a popular parallel processing tool in the era of big data. While practitioners have rewritten many conventional analysis algorithms to make them customized to Hadoop, the issue of inefficient I/O in Hadoop-based programs has been repeatedly reported in the literature. In this article, we address the problem of the I/O inefficiency in Hadoop-based massive data analysis by introducing our efficient modification of Hadoop. We first incorporate a columnar data layout into the conventional Hadoop framework, without any modification of the Hadoop internals. We also provide Hadoop with indexing capability to save a huge amount of I/O while processing not only selection predicates but also star-join queries that are often used in many analysis tasks.