Summary

  • An Internet Society Pulse guest article published on 22 September celebrated 54.27% dual-stack coverage in IPv6Matrix's August host crawl, while its performance comparison drew on an older May snapshot.
  • IPv6Matrix's public API now includes an August paired-ping row: mean response times were about 343.1 ms over IPv4 and 93.1 ms over IPv6, but IPv6 was lower for 324,957 of 2,504,699 paired hosts—12.97%.
  • The mean, the paired-host count, the crawl date and the crawler's domain-derived population answer different questions; none alone describes every user's experience or proves an operator's service quality.

The attractive number in an IPv6 migration presentation is 93 milliseconds. IPv6Matrix's current August ping series produces that mean for IPv6 against roughly 343 milliseconds for IPv4. Both are computed from the sums of responses for 2,504,699 successfully pinged dual-stack hosts. If the slide stops there, it leaves out another field published by the same project: faster, defined in its API documentation as the number of hosts with a lower IPv6 than IPv4 response time. In the August row it is 324,957, or 12.97% of the paired population.

That is not a paradox and it is not evidence that the data are broken. A mean can be pulled by an extreme tail even when comparatively few individual pairs favour the lower-mean side. The public summary does not give a median, a distribution of per-host differences or the handling of ties, so the other 87.03% must not be casually renamed “IPv4 wins.” The defensible statement is narrower: the August aggregate IPv6 mean is lower; a much smaller counted share of these paired host observations has lower IPv6 latency. Neither number specifies the latency a particular subscriber, application or region will see.

The dates also matter. Olivier Crépin-Leblond's 22 September Internet Society Pulse guest article reports the 3 August host crawl's 2,897,950 dual-stack hosts among 5,340,321, or 54.27%. For performance it cites May means of 333.0 and 93.8 milliseconds. The fuller Internet Society UK Chapter report explicitly labels its supporting ping and reachability series as May while the host headline is August; its May paired-ping table already counted only 308,646 of 2,529,441 hosts, 12.20%, as lower-latency on IPv6. It also warns about a long tail. This is not a discovery that the report hid its denominator.

It is a reason not to detach a performance number from the vintage and paired count the report supplied.

Today IPv6Matrix's own public API contains an August ping entry and an August reachability entry. The API response inspected for this article did not state when those rows became available. It would be unwarranted to say that the report ignored an August performance series already published on its issue date. It is enough to say that a reader using the live feed now sees a later companion observation than the one frozen in the report.

Even a second headline can mislead if its unit changes. The current August reachability series counts 3,855,760 responses among 4,338,714 IPv4 service hosts, 88.87%, and 1,547,097 among 1,618,920 IPv6 service hosts, 95.56%. Those are protocol-specific service populations, not a paired trial in which the same hosts failed on one protocol and succeeded on the other. IPv6Matrix builds its population by starting with DNS records for the Majestic Million domains and identifying their web, mail, time and name-service hosts. It is not a traffic-weighted census of users.

The practical discipline is simple: attach a measurement receipt whenever an adoption or performance statistic is copied into a policy paper, purchasing criterion or network promise. Record the crawl date, population, unit, denominator, field definition, means and paired-host share; say whether the citation is a fixed report or today's API. This is Daniel Kade's editorial recommendation, not a rule adopted by the project or Internet Society. The August data support real dual-stack growth. They do not license the slogan that IPv6 is faster for almost everyone.

Sources