[tsvwg] ECN packet counters from a Czech ISP

[Pete Heist <[e-mail]>]

There's been some discussion starting around how much ECN negotiating
and marking is going on on the Internet. Although we know of a number
of vendor products that ship with fq_codel and mark by default, and
following the precautionary principle seems appropriate with congestion
control signals, beyond just vendor lists and anecdotal experience it
would also help to try to look at some ECN stats, globally if possible.

Earlier this year I was able to convince the cooperative ISP I belong
to in Czech to gather some simple ECN counter stats at the gateway with
iptables. We ran counters for around 30TB of traffic over 11 days. I've
finally taken the time to post those results:

https://github.com/heistp/ecn-counters

We intend to write up a further analysis of these results and what the
implications are for our recent l4s-tests batch, but I see no harm in
posting the data first.

Caution #1: As warned in the repo, it's always possible that something
is wrong with the way this data was collected or how the packet matches
work in this particular environment. However, the results appear sane,
he said hopefully.

Caution #2: Before seeing a low number or percentage and drawing a
conclusion, make sure to understand what those numbers mean. Also,
there are also some limitations to the stateless counter approach, but
I still think there's some useful info here as well.

Pete