The vast majority of popular Internet applications, from electronic mail to Web search, and from video streaming to multi-player games, run on servers managed by Online Service Providers (OSPs) like Google, Microsoft, or Akamai.  In this project, we: (1) design and evaluate techniques for OSPs to offer good performance to users at reasonable cost, by optimizing how they create, store, and deliver content; (2) study how both access link speeds and the utilization/congestion of the interconnects between ISPs and OSPs ultimately affect application performance. Our research ranges from measuring and characterizing existing services to designing and evaluating new measurement and optimization techniques, often by collaborating with operational OSPs or the Internet service providers that help them deliver traffic to users.  

 The main outcomes of the project are:

 Measurement-based diagnosis of OSP performance: Our measurement studies have explored the current state of utilization at Internet interconnection points and showed that, contrary to conventional public wisdom, these interconnection points typically have ample spare capacity to support applications such as streaming video. We have uncovered the root cause of performance problems for video streaming (jointly with Yahoo) causes of congestion at Internet eXchange Points. We have also studied how wireless network performance in the home can also prove to be a limiting factor for fixed and mobile broadband Internet access; these results have formed the core of reports published by the Federal Communications Commission and others. These studies have also allowed researchers and practitioners to identify and fix underlying performance problems, and have informed the broader policy debate on the challenges of how the Internet supports the delivery of high-bandwidth video content.

 Optimizing the performance of OSPs: We have designed and evaluated new techniques for OSPs to optimize network paths and sending rates within private backbone networks (jointly with Google) and data-center networks (jointly with Barefoot Networks and VMware), the network topology and routing in wide-area networks (jointly with Sodero Networks and AT&T), allocation of disk space in edge caching servers (jointly with Akamai), and server selection and routing in delivering data across the Internet to end users. Our PECAN work showed that traffic engineering offers minimal benefit beyond optimal cache placement for improving the performance of Web applications. Our work studying the relationship of OSP application performance and access network throughput showed that certain applications, such as Web page loads, suffer diminishing returns as access link performance increases; the optimizations that we implemented in home routers that prefetch the DNS records of commonly requested web pages and maintain active TCP connections to popular web pages can reduce page load time for popular sites in the home while incurring only minimal data usage overhead.

 Network mechanisms to improve performance: We have designed new mechanisms for network devices, such as home routers or data-center switches, to help improve performance. For example, we created new measurement techniques that can collect performance data for individual users efficiently on high-speed switches, a new programmable software switch to enable innovation in data-center networks, and a flexible network assembly language for programming future high-speed switches.

Together, these new techniques give OSPs a much better way to measure, optimize, and design their services -- and the underlying network -- to offer good performance at reasonable cost.  In addition, the project led to the training of numerous undergraduate and graduate students in networking research, as well as the creation of open-source software for the broader community.

Last Modified: 06/16/2017
Modified by: Jennifer Rexford