Making the Internet Fast

Ethan Katz-Bassett, assistant professor in the computer science department at the University of Southern California

Video of lecture

Abstract

Ethan Katz-Bassett

For web services, time is money. For example, Amazon found that every additional 100 ms of delay in loading a page costs them 1% of their sales. Despite this need for speed, Internet performance often lags.

A more robust Internet–one with the predictable performance needed to provide critical services–requires the development of a new generation of better tools and protocols. In this talk, I will present three pieces of work we’ve done on improving Internet performance for large web services. The work involves deployable solutions that can benefit applications today and enable applications for tomorrow.

First, I will present Reverse Traceroute, our system to measure reverse paths back to the local host from other networks. While tools have long existed to measure the forward direction, the reverse path has been largely opaque, hindering troubleshooting efforts. I will show how Google and other content providers can use Reverse Traceroute to troubleshoot their clients’ performance problems.

Second, I will present three novel TCP loss recovery mechanisms that we developed in collaboration with Google. Proactive, Reactive, and Corrective are three qualitatively-different, easily-deployable mechanisms that (1) proactively avoid losses, (2) recover from them as quickly as possible, and (3) reconstruct packets to mask loss. Our large-scale experiments on Google’s production network demonstrate a 23% average decrease in Google client latency.

Finally, I will present results from our measurement study that exposed a massive change in strategy and expansion of Google’s serving infrastructure. We developed novel techniques to locate all Google servers around the world, as well as uncover the client-to-server mappings, and happened to discover a 7x increase in the number of Google server locations around the world. I will present our preliminary results on understanding the performance impact, limitations, and potential of this huge expansion.

Bio

Ethan Katz-Bassett is an assistant professor in the computer science department at the University of Southern California, where he works on systems and protocols to dramatically improve Internet reliability and performance. His research is funded by an NSF CAREER Award, Google Faculty Research Awards, and an M-Lab Network Research Grant. He runs a networking and systems research group with Ramesh Govindan, Minlan Yu, and Wyatt Lloyd. Prior to joining USC, Ethan worked at Google on mobile web performance. He received his PhD from the University of Washington.