IPv6 is Faster

A couple of weeks ago, I ran a panel at NANOG60 about the IPv6 Performance Bonus.

First, Geoff Huston showed some results from his experiments. He buys Google ads which run a bit of code, telling the browser to fetch a v6-only object and a v4-only object, and compares the times. He notes that in IPv4, a host will send a SYN, but never respond to the SYN-ACK with an ACK, 0.1-0.3% of the time. With IPv6, the connection fails 1.2-2.1% of the time. He speculates that this is due to IPv6-incapable filters or firewalls. Others have speculated that state machines, such as DPI or proxies, may not have robust IPv6 implementations yet. But we don't really know.

That's a recurring theme in this presentation. Geoff notes that his IPv4:IPv6 comparison showed, at World IPv6 Launch, that IPv4 was a bit faster on average. However, in December 2013, he shows that many samples show IPv6 faster, by a ration of 1.1-1.2 to 1; just 10-20% faster. He speculates that they might use different paths, or less middleware (etc.) in IPv6. But we don't really know.

I reported on data seen at Time Warner Cable, using hub-based servers to measure round-trip time (RTT) to top web sites over IPv4 and IPv6. Although IPv6 was not faster from all locations to all sites, the average IPv6 RTT was 15% better than the average IPv4 RTT. Trying to figure out why, I show a comparison of traceroute hop count, and IPv6 on average does have one less hop than IPv4. There doesn't seem to be a strong correlation between the different hop count and RTT between families; some web sites or transit paths may be tunneled. But we don't really know.

It's operationally useful to know when the RTT and hop count changed; one could then correlate with network events and look for causes. But nobody has.

John Brzozowski talked about Comcast's measurements, using a variety of sources to feed a common store. One specific data set is the use of dual-stacked speed tst servers, which shows IPv6 faster than IPv4. He said this may be related to CDN support, more recent chipsets being better, and more optimized MTU. But we don't really know.

The weirdest report came from Chris Donley, who did some lab testing showing IPv4 via NAT being significantly slower than IPv6, but when he put another home NAT in line, the speed difference vanished. This would probably not be borne out in a CGN scenario, since IPv6 traffic would not use the CGN (whether NAT44 or DS-Lite). But we don't really know.

To recap:

  • Geoff Huston's tests from browser to a web server show IPv6 faster.
  • Lee Howard's tests from hub to web sites shows IPv6 faster.
  • John Brzozowski's tests from browser to speed test server show IPv6 faster.
  • Chris Donley's lab test through home NAT shows IPv6 faster.

As often happens when you ask an interesting question, more questions are raised than answered. We don't know why IPv6 is faster than IPv4, but every current measurement says that it is.

Lee Howard
2 March 2014