Building Autonomic Elements from Video-Streaming Servers

HTTP Streaming is nowadays the main approach for delivering video-streaming on the Internet. As a consequence of that, the widely deployed HTTP infrastructures face new challenges posed by the sensitivity of video-streaming users to service quality degradation and the specificities of video-streaming workloads. Performance issues represent one main class of problems in the server infrastructure that can result into a significant deterioration of the end-users’ quality of experience (QoE), proportional to the upfront time spent by them watching the videos. This paper addresses the development of autonomic HTTP Streaming servers organized into Autonomic Elements (AEs), the building blocks of Autonomic Computing (AC) systems. AEs are structured using container-based virtualization and are provided with monitoring, failure prediction, failure diagnosis and repair features. These features are incorporated into SHStream, a self-healing framework developed by us. SHStream relies on online learning algorithms to build and evaluate classification models dynamically for prediction and diagnosis of performance anomalies. The results of our experimental analysis have shown that: (1) failure prediction can be performed with approximately 98%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$98\%$$\end{document} of recall and 99%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$99\%$$\end{document} of precision; (2) the diagnosis activity can localize and identify the resource responsible for performance failures, without misclassifications; (3) the classifiers’ performance stabilizes using a small number of learning instances; and (4) container-based virtualization technologies enable recovery times shorter than 1 s through rebooting and shorter than 3 s using server migration techniques.