An empirical approach to automated performance management for elastic n-tier applications in computing clouds.

摘要

Achieving a high degree of efficiency is non-trivial when managing the performance of large web-facing applications such as e-commerce websites and social networks. While computing clouds have been touted as a good solution for elastic applications, many significant technological challenges still have to be addressed in order to leverage the full potential of this new computing paradigm. In this dissertation I argue that the automation of elastic n-tier application performance management in computing clouds presents novel challenges to classical system performance management methodology that can be successfully addressed through a systematic empirical approach. I present strong evidence in support of my thesis in a framework of three incremental building blocks: Experimental Analysis of Elastic System Scalability and Consolidation, Modeling and Detection of Non-trivial Performance Phenomena in Elastic Systems, and Automated Control and Configuration Planning of Elastic Systems. More concretely, I first provide a proof of concept for the feasibility of large-scale experimental database system performance analyses, and illustrate several complex performance phenomena based on the gathered scalability and consolidation data. Second, I extend these initial results to a proof of concept for automating bottleneck detection based on statistical analysis and an abstract definition of multi-bottlenecks. Third, I build a performance control system that manages elastic n-tier applications efficiently with respect to complex performance phenomena such as multi-bottlenecks. This control system provides a proof of concept for automated online performance management based on empirical data.