[v6ops] Re: We should document IPv6 packet loss...

[Brian Candler <[e-mail]>]

>> Mean IPv6 latency: 294 ms
>> Mean IPv4 latency: 264 ms

Another minor comment: arithmetic mean is heavily influenced by large 
outliers, which means those numbers aren't easy to interpret.

Median and 90th percentile would be interesting to see.  Possibly 
harmonic mean too (since RTT is unlikely to be significantly *lower* 
than the minimum possible)

Since there are three tests to each target, some indication of how big 
the spread is between the shortest and longest would be interesting too.


>> IPv6 success rate: 90.3 %
>> IPv4 success rate: 98.7 %

Does "success" count each individual test, so that one particular target 
could count as 0, 1, 2 or 3 towards the total?

I think it would be interesting to see success/fail divided up into 
three categories per target:

(a) all three probes succeeded
(b) one or two probes succeeded
(c) zero probes succeeded

Case (b) would be odd to see, because TCP should retransmit, and simple 
packet loss should show itself as longer time to establish connection. 
But it's worth knowing if it happens at all.

That leaves all remaining failures as case (c). That's not what I think 
of as "packet loss" though. It probably means either:

- the probe has no IPv6 connectivity at all, even though it advertises 
that it has

- there's an Internet topology issue between the probe and the test 
location, i.e. incomplete routing tables somewhere. Peering politics 
between certain providers can cause that. Or a firewall with different 
policies for v4 and v6.

Regards,

Brian.