Performance analysis using POP methodology in spark big data applications

• Performanceanalyse von Big-Data-Anwendungen in Spark unter Nutzung der POP-Methodik

Brückner, Moritz; Müller, Matthias S. (Thesis advisor); Geisler, Sandra (Thesis advisor); Liem, Radita Tapaning Hesti (Consultant)

Aachen : RWTH Aachen University (2023)
Bachelor Thesis

Bachelorarbeit, RWTH Aachen University, 2023

Abstract

Today’s software applications need to cope with ever increasing amounts of data while processing the data in a reasonable amount of time with limited resources. Specialized frameworks such as Apache Hadoop or Apache Spark are often used to meet those requirements, making it possible to run an application in a distributed and parallel manner on multiple compute nodes in a cluster network. A common issue with these frameworks is that both the configuration of an applications as well as the kind of application and the structure of its data are strongly influencing the application’s performance. In addition to that, there seems to be a current trend of convergence of the originally largely independent disciplines of high-performance computing (HPC) and big data, whose applications increasingly overlap. As a result, the application of Apache Spark on HPC systems is gaining relevance and, consequently, also the study of performance of Spark applications on these systems. In this thesis, the POP methodology, originally developed for analyzing the performance of HPC applications, is applied to Spark big data applications. The core principle of the methodology is to assign a score to individual performance-influencing aspects, which can be used to obtain a comprehensive and direct overview of potential performance bottlenecks of an application. The aim of this thesis is to evaluate selected Spark benchmarks from the HiBench benchmark suite and to use the obtained results to derive POP metrics for Spark applications. Beyond the POP metrics that are used in the HPC context, additional Spark-specific metrics are proposed in order to significantly extend the range of identifiable problems and to allow for a more precise determination of these problems. This thesis comes to the conclusion that, in principle, the POP methodology can be successfully applied to Spark applications, although in some cases certain limitations or assumptions are necessary. Even though it is not possible to verify the correctness and completeness of the proposed metrics beyond any doubt by means of the conducted experiments, the methodology presented in this thesis seems to be suitable for identifying a large number of different performance problems. Yet, further investigations are required in order to eliminate some of the limitations and assumptions made, and to improve and validate both individual metrics as well as the methodology as a whole.

Institutions

• IT Center [022000]
• Department of Computer Science [120000]
• Chair of Computer Science 12 (High Performance Computing) [123010]