NeTS: Small: Data Driven Mobile Web Performance

Brown University, Providence RI

Investigators

Abstract

The world is growing increasingly dependent on web services; they are now a vital part of business, personal and social aspects of daily life. Unfortunately, users continue to be plagued by a myriad of performance problems, e.g., videos that stall while buffering, slow web page load times, and web pages that crash. These performance problems occur for multiple reasons, the principal one being a lack of correctly configured and adequately tuned networking protocols and algorithms that can deliver the desired quality of service. Moreover, as the set of web services and technologies continues to grow and broaden, the performance implications of these poorly configured networks are likely to become more widespread and apparent. The project vision is to develop an alternative approach to managing and configuring web servers that can meet end user's performance expectations: a method which improves performance by tailoring the web server's configurations to match the specifics of each application and end user's technology. If successful, this proposal will fundamentally impact widely accepted principles for managing services and will enable a broad range of users to benefit from efficient operation of the growing set of essential web services. This project proposes an ambitious research agenda based on a data-driven specialization of the web serving stacks, wherein the stack learns the optimal protocol optimizations and parameters for each user and, in turn, uses this information to specialize or tune the web stack to improve the end user's experience. The proposed research has five thrusts: First, the project will conduct a large-scale systematic study on the interactions between protocol configuration options, e.g., optimizations, device profiles, web complexity, and network conditions and develop tools to analyze and interpret the implications of these interactions. Second, it will develop a new architecture for managing web servers that provides a flexible and uniform interface for collecting telemetry data and configuring protocol parameters and optimizations. Third, the researchers will develop a dynamic learning framework that scalably and accurately learns and predicts the optimal configuration for each user. Fourth, the project proposes to develop a novel programming interface that enables end-users and their devices to participate in the learning process by providing feedback and meta-data that is only available on the client side. Finally, the researchers will consider a number of design and implementation challenges necessary to realize this new approach, including adjusting to drastic changes in network conditions, tackling complex relationships between configuration parameters and user experience, making real-time predictions, etc. This project will actively undertake tasks to realize broader impact through educational outreach across all levels of education, with a focus on undergraduate and high school students, and under-represented minorities in Science, Technology, Engineering, and Mathematics. 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.