Evaluating Hadoop Clusters with TPCx-HS

TL;DR

This paper evaluates how network configurations affect Hadoop cluster performance using the TPCx-HS benchmark, revealing significant performance gains with dedicated networks despite similar costs.

Contribution

It provides an experimental comparison of shared versus dedicated network setups for Hadoop clusters using the TPCx-HS benchmark.

Findings

Dedicated networks outperform shared networks in Hadoop performance.

Performance gains justify the minimal additional cost.

TPCx-HS effectively stresses network impact on Hadoop clusters.

Abstract

The growing complexity and variety of Big Data platforms makes it both difficult and time consuming for all system users to properly setup and operate the systems. Another challenge is to compare the platforms in order to choose the most appropriate one for a particular application. All these factors motivate the need for a standardized Big Data benchmark that can help the users in the process of platform evaluation. Just recently TPCx-HS [1][2] has been released as the first standardized Big Data benchmark designed to stress test a Hadoop cluster. The goal of this study is to evaluate and compare how the network setup influences the performance of a Hadoop cluster. In particular, experiments were performed using shared and dedicated 1Gbit networks utilized by the same Cloudera Hadoop Distribution (CDH) cluster setup. The TPCx-HS benchmark, which is very network intensive, was used to stress test and compare both cluster setups. All the presented results are obtained by using the officially available version [1] of the benchmark, but they are not comparable with the officially reported results and are meant as an experimental evaluation, not audited by any external organization. As expected the dedicated 1Gbit network setup performed much faster than the shared 1Gbit setup. However, what was surprising is the negligible price difference between both cluster setups, which pays off with a multifold performance return.