Performance Prediction for Enabling Intelligent Resource Management on Big Data Processing Workflows

• Main Author: Obeso Duque, Aleksandra
• Format: Others
• Language: English
• Published: Uppsala universitet, Institutionen för informationsteknologi 2018
• Subjects: Engineering and Technology; Teknik och teknologier
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-372178

Mobile cloud computing offers an augmented infrastructure that allows resource-constrained devices to use remote computational resources as an enabler for highly intensive computation, thus improving end users experience. Being able to efficiently manage cloud elasticity represents a big challenge for dynamic resource scaling on-demand. In this sense, the development of intelligent tools that could ease the understanding of the behavior of a highly dynamic system and to detect resource bottlenecks given certain service level constrains represents an interesting case of study. In this project, a comparative study has been carried out for different distributed services taking into account the tools that are available for load generation, benchmarking and sensing of key performance indicators. Based on that, the big data processing framework Hadoop Mapreduce, has been deployed as a virtualized service on top of a distributed environment. Experiments for different cluster setups using different benchmarks have been conducted on this testbed in order to collect traces for both resource usage statistics at the infrastructure level and performance metrics at the platform level. Different machine learning approaches have been applied on the collected traces, thus generating prediction and classification models whose performance is then evaluated and compared. The highly accurate results, namely a Normalized Mean Absolute Error below 10.3% for the regressor and an accuracy score above 99.9% for the classifier, show the feasibility of the prediction models generated for service performance prediction and resource bottleneck detection that could be further used to trigger auto-scaling processes on cloud environments under dynamic loads in order to fulfill service level requirements.