IPv6 and CDN's

Dave Taht dave.taht at gmail.com
Sun Nov 28 23:06:17 UTC 2021


On Sat, Nov 27, 2021 at 12:18 PM William Herrin <bill at herrin.us> wrote:
>
> On Fri, Nov 26, 2021 at 3:07 PM Michael Thomas <mike at mtcc.com> wrote:
> >> On 11/26/21 1:44 PM, Jean St-Laurent via NANOG wrote:
> >> Here are some maths and 1 argument kicking ass pitch for CFO’s that use iphones.
> >> Apple tells app devs to use IPv6 as it's 1.4 times faster than IPv4
>
> > This really hits my bs meter big time.
>
> If I had to guess, this is an example of correlation is not causation.
> Folks with IPv6 tend to be on savvier service providers who have
> better performance for both IPv4 and IPv6.  To find out for sure,
> you'd have to do an experiment where same-user-same-server connections
> are split between IPv4 and IPv6 and then measure the performance
> difference. I don't know if anyone has done that but these particular
> articles look like someone is just looking at the high-level metrics.
> Those won't hold any statistical validity because they're not actually
> random samples.

We wrote 110+ tests for flent.org to test services under load, you can
use -4 or -6, and all the plots against all different sorts of test
conditions, can be compared against each other easily.

Example of use:

flent --socket-stats --step-size=.05 -l 300 -H
fremont.starlink.taht.net -4 -t ipv4 rrul
flent  --socket-stats --step-size=.05 -l 300 -H
fremont.starlink.taht.net -6-t ipv6 rrul
flent-gui *.flent.gz # and add-other-datafiles

There are also several tests in the flent suite (like rrul46) that try
ipv4 and 6 at exactly the same time. You typically run into (ISP)
bottleneck or (DC) host bandwidth limits first, unless you also
distribute the generated load across multiple machines (I use pdsh for
this, I'd like to know of tools in modern clouds to fire off a load
generator like this, or of other load generators). We have also been
using the irtt tool at a very high resolution (3ms) to map networks'
jitter and latency at "idle" with rather interesting results for
3/4/5g, wifi and starlink, but haven't been breaking down that data
between ipv4 and ipv6 as yet.

Other useful statistics to perhaps gather at scale would be TCP_INFO
rtt, loss, marking, retransmit stats from customer-facing apache or
other proxy servers, and break those down between ipv4 and ipv6 and by
AS.

> Regards,
> Bill Herrin
>
>
> --
> William Herrin
> bill at herrin.us
> https://bill.herrin.us/



-- 
I tried to build a better future, a few times:
https://wayforward.archive.org/?site=https%3A%2F%2Fwww.icei.org

Dave Täht CEO, TekLibre, LLC


More information about the NANOG mailing list