Re: IPv6 Performance (was Re: IPv4 Pricing)

[Marco Moock via NANOG <nanog () lists nanog org>]

Am 04.12.2025 um 06:22:19 Uhr schrieb Vasilenko Eduard via NANOG:

> Yes, indeed, https://stats.labs.apnic.net/v6perf shows -6.1ms
> globally. Highly probably, Geoff did everything right, we could trust
> the numbers. But sorry, there is no explanation why?

There are various reasons:

No NAT/CGNAT
No DS-Lite tunneling (less MTU and more CPU cycles on the CPE chips)
Smaller routing tables
Different peering/routing inside AS

If you really want the proof, you have to choose one ASN and examine
all the devices and check the latencies there.
There is no "this is the general reason for all". Some routers also
handle SPI/NAT/routing different, especially when longer prefixes are
used.
I heard (never verified) that certain devices only handle IPv4 /24 and
IPv6 /48 fast, the rest will be processed slower due to ASIC chip
design.

> At least, I have never seen anything that could be called "prove".
> Your speculation is the same reliable as mine. Actually, mine (about
> new networks -> high quality) looks more probable for me.

Explain the term "new". Various ISPs exist for decades and I have
serious doubt that old devices are still present there. Even entire
network architectures were outphased, e.g. ATM, ISDN and old DSL
infrastructure.

> 2.
> RTT does not matter. Not at all.

That is indeed right. I saw situations where the ICMP messages of a
router more hops away arrived faster than one from a hop closer. A
reason is also that this stuff is most likely not processed in the ASIC
based forwarding plane, but in a slower control plane. If you want to
know more about that, you have to ask the router's manufacturers.

> Because it is not visible for the
> end user directly and it has negligible influence for the good
> Congestion Control (like BBR). User cares only about the FCT (flow
> completion time). FCT is dependent on (in priority): 1) bottleneck
> bandwidth, 2) packet loss, 3) congestion control convergence, 4) RTT
> (but only if congestion control is bad), and a few other things.
> Subtracting 2.6(6)% of the goodput (because of bigger overhead) make
> IPv6 fundamentally slower (for the all other things equal).

You said that so many times, it is boring now and that impact doesn't
have enough weight to change the average latency in a way that IPv6 is
slower in general.

-- 
Gruß
Marco

Send unsolicited bulk mail to 1764825739muell () cartoonies org

Attachment: _bin
Description: Digitale Signatur von OpenPGP

_______________________________________________
NANOG mailing list 
https://lists.nanog.org/archives/list/nanog () lists nanog org/message/27M4B5AKFMNH65AAPX2KQMUZADKHG3ZQ/