Evaluation of Apache Hadoop for parallel data analysis with ROOT

Abstract

The Apache Hadoop software is a Java based framework for distributed processing of large data sets across clusters of computers, using the Hadoop file system (HDFS) for data storage and backup and MapReduce as a processing platform. Hadoop is primarily designed for processing large textual data sets which can be processed in arbitrary chunks, and must be adapted to the use case of processing binary data files which cannot be split automatically. However, Hadoop offers attractive features in terms of fault tolerance, task supervision and control, multi-user functionality and job management. For this reason, we evaluated Apache Hadoop as an alternative approach to PROOF for ROOT data analysis. Two alternatives in distributing analysis data were discussed: either the data was stored in HDFS and processed with MapReduce, or the data was accessed via a standard Grid storage system (dCache Tier-2) and MapReduce was used only as execution back-end.The focus in the measurements were on the one hand to safely store analysis data on HDFS with reasonable data rates and on the other hand to process data fast and reliably with MapReduce. In the evaluation of the HDFS, read/write data rates from local Hadoop cluster have been measured and compared to standard data rates from the local NFS installation. In the evaluation of MapReduce, realistic ROOT analyses have been used and event rates have been compared to PROOF.