Reducing short flows' latency in the internet

Short flows are highly valuable in the modern Internet and are widely used by applications in the form of web requests or with user interactions. These kinds of applications are extremely sensitive to latency. A small additional delay, like one or two round trip times (RTTs), may easily cause user frustration and lose usability of services. In the most desirable scenario, we want to finish these kinds of flows in one network RTT. Furthermore, we would like the network's RTT to be as close as possible to the speed of light. Unfortunately, in the current Internet, there are many unnecessary delays caused by different kinds of policies--in particular, transmission protocol and routing policies--driving us far away from this goal. This thesis aims at answering the following two questions: How can we optimize the transmission protocol to reduce short flows' latency as close as possible to one RTT and why are network RTTs still significantly larger than the speed-of-light latency? To reduce the transmission latency, we focused on the two main components of short flows, connection establishment and data transmission. ASAP, a new naming and transport protocol, is introduced to reduce the time spent on initial TCP connections. It merges functionality of DNS and TCP's connection establishment functions by piggybacking the connection establishment procedure atop the DNS lookup process. With the help of ASAP, the host is able to save up to two-thirds of the time spent on initial connection without exposing significant DoS vulnerabilities. For data transmission, we designed a new control rate mechanism, Halfback, which achieves low latency with limited bandwidth overhead and only requires sender-side changes. Halfback has an aggressive startup phase, finishing transmission for most short flows in one RTT, together with a Reverse-Ordering Proactively Retransmission phase which helps the host to recovery quickly from packet loss caused by the aggressive startup phase. Halfback is able to achieve 56% smaller flow completion time on average and three times smaller in the 99th percentile. RTT between two hosts is able to be more than 6 times the speed-of-light latency for Directed Optical Fiber. To understand the composition of RTT inflation, we break down the path inflation on the end-to-end path into its contribution factors. Based on our result, 7.2% is caused by network topology, 18.8% is contributed by inter-domain routing policies, 54.9% is caused by peering policies, and 25.6% is caused by intra-domain routing policies. This result shows that the main component of the path inflation is caused by peering policies which may require more attention for future research. Besides this, we also analyze the changes of the inflation caused by each contributing factor across five years. According to our analysis, the total inflation has been reduced by around 6% each year since 2010.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2016

Author dc:creator

• Li, Qingxi

Contributors dc:contributor

• Godfrey, Phillip Brighten
• Caesar, Matthew
• Gupta, Indranil
• Taft, Nina

• Transmission control protocol (TCP)
• Short flows
• Flow completion time

Statement dc:rights

• Copyright 2015 Qingxi Li

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/90521

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/90521