Slow going for TCP Fast Open

As part of our continued effort to measure Internet path transparency with PathSpider, we’ve taken a look at the state of deployment and potential impairments to TCP Fast Open. TCP Fast Open is an extension to TCP that allows data to be placed on the first (SYN) packet of the TCP handshake, eliminating a round-trip time from TCP connections. It uses a TCP Option to exchange a cookie to be used on subsequent fast open connection attempts, to reduce the risk of TFO-based denial of service attacks.Interference with this option could cause path impairment of TFO, and indeed Christoph Paasch has reported that this is the case on about 20% of the access networks he observed.

We set out to measure possible impairment on content provision networks and in the Internet core, and found instead that TFO deployment on popular Web servers is mostly limited to Google, who invented TFO. Of 939,680 web servers taken from the Public Targets List (PTL), only 866 (0.092%) negotiated TFO in measurements taken this week. 690 (about 79.7%) of these are Google servers. Compared to measurements taken in October 2016, we see no appreciable change; then 563 of 635,681 web servers (0.086%) negotiated TFO. This is unsurprising, given that TFO requires significant changes to both client-side and server-side application logic as well as kernel support on both endpoints, we expect slow adoption compared, e.g., to ECN.

The story on DNS, where TFO is part of an effort to improve query privacy by using TLS and TCP for DNS, is similar: of the 53,267 authoritative name servers taken from the PTL, 56 (0.105\%) negotiate TFO, only three of which are not Google name servers; two of those three use an experimental ID, and fail to ACK data on the SYN.

About Brian Trammell

Brian Trammell is an Internet measurement and architecture geek, a senior researcher at ETH Zürich's Networked Systems Group, and a member of the Internet Architecture Board.