CAREER: Enabling Predictable Performance in Cloud Computing

点击翻译按钮获取中文摘要

Cloud computing allows tenants, such as Netflix and Expedia to economically rent compute and storage resources from providers. To enable low resource prices, providers consolidate multiple tenants onto a single physical server. However, this sharing of physical resources among tenants often leads to contention, resulting in unpredictable performance. Worse, tenants cannot observe resource contention due to the opaque nature of cloud computing. This project will develop novel performance models to estimate resource contention in opaque cloud deployments. These models will then be leveraged to develop solutions for cloud tenants that mitigate performance variation, thus enabling predictable performance in clouds.To realize predictable performance, the project will proceed along two integrated fronts. On the theoretical front, the project will develop uncertainty-aware stochastic performance models. These models will then be integrated with control-theoretic and machine learning techniques to infer, at runtime, the unobservable model parameters in a cloud environment. On the systems front, armed with the uncertainty-aware models, the project will develop solutions, including task schedulers and resource managers, that alleviate application performance variation. The solutions will be designed to dynamically detect and diagnose performance interference. All models and solutions will be experimentally evaluated in public and private clouds.The interdisciplinary nature of the project provides unique opportunities for integrated education and outreach. The primary benefit of the project will be increasing cloud adoption and promoting its broader impact on energy efficiency. To facilitate this goal, the project will develop open-source solutions for platforms such as OpenStack. The project will advance interdisciplinary education by developing performance analysis lectures and modules that will be integrated with existing courses taught in the departments of Computer Science and Applied Mathematics and Statistics, and the College of Business. Outreach activities will focus on creating research opportunities for local area high school students.All data produced as a result of this project, including models, software solutions, publications, and courseware, will be made publicly available at the project repository: http://www.pace.cs.stonybrook.edu/predictable-clouds.html. The project data will be maintained and made available for at least 10 years, and even longer, if needed. Data will be stored and hosted on local web servers, and will also be replicated on external public web servers, such as those provided by github, which offer long-term durability and reliability.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.