Evaluation and optimization of Big Data Processing on High Performance Computing Systems

Resumo

Nowadays, Big Data technologies are used by many organizations to extract valuable information from large-scale datasets. As the size of these datasets increases, meeting the huge performance requirements of data processing applications becomes more challenging. This Thesis focuses on evaluating and optimizing these applications by proposing two new tools, namely BDEv and Flame-MR. On the one hand, BDEv allows to thoroughly assess the behavior of widespread Big Data processing frameworks such as Hadoop, Spark and Flink. It manages the configuration and deployment of the frameworks, generating the input datasets and launching the workloads specified by the user. During each workload, it automatically extracts several evaluation metrics that include performance, resource utilization, energy efficiency and microarchitectural behavior. On the other hand, Flame-MR optimizes the performance of existing Hadoop MapReduce applications. Its overall design is based on an event-driven architecture that improves the efficiency of the system resources by pipelining data movements and computation. Moreover, it avoids redundant memory copies present in Hadoop, while also using efficient sort and merge algorithms for data processing. Flame-MR replaces the underlying MapReduce data processing engine in a transparent way and thus the source code of existing applications does not require to be modified. The performance benefits provided by Flame- MR have been thoroughly evaluated on cluster and cloud systems by using both standard benchmarks and real-world applications, showing reductions in execution time that range from 40% to 90%. This Thesis provides Big Data users with powerful tools to analyze and understand the behavior of data processing frameworks and reduce the execution time of the applications without requiring expert knowledge.