Over-utilisation of v6 neighbour slots

[Ralph Droms (rdroms)]

On Nov 1, 2013, at 3:46 PM 11/1/13, Erik Kline <ek at google.com> wrote:

> I'm not sure what you mean by "impact".

Sorry for the lack of clarity.  You answered my question - the different behaviors are strictly correlated to the version of OS 10 and are independent of the version of Safari.

- Ralph

> 
> Different OS versions do seem to behave differently.  I think 10.7 is
> when the dueling connections implementation was introduced, and it did
> drop Safari's IPv6 native preference by basically half.  And if you
> think about it, in a perfect network where IPv4 and IPv6 have
> identical performance each would likely only be used ~50% of the time.
> 
> It looks like you would basically need to penalize IPv4 in order get
> your IPv6 ROI, *if* all your customers ever do is surf the web using
> Safari on recent Macs to do so.  The total impact of an HE
> implementation is obviously in proportion to its use for web traffic.
> 
> Let me revise the numbers for this sample network.  I was including
> measurements where the client may not have asked for any AAAAs, so let
> me restrict my numbers just to measurements where AAAAs were
> requested.  That's what I get for not waiting to have my stats-diving
> be peer reviewed (different timezone, etc).
> 
> "Once more, with feeling!", and I'm sure Lorenzo and correct me later on.
> 
> [All MSIE]
>    no HE, later versions do a simple reachability check ~once
>    96% IPv6 native preference
>    7ms average extra latency
> 
> [All Chrome]
>    HE: 300ms lead for IPv6 then fallback to IPv4
>    83% IPv6 native preference
>    1.5ms average latency improvement over IPv4
> 
> [All Safari + 10.6]
>    I think no HE?  (it's been so long...)
>    97.5% IPv6 native preference
>    0ms average latency delta between IPv4 and IPv6
> 
> [All Safari + 10.7,10.8, and 10.9]
>    HE: strict connection racing, at one point A record requested
> first, I believe
>    48% IPv6 native preference
>    9ms average latency improvement over IPv4 (by design, working as
> intended, etc.)
> 
> So the HE implementation does matter.  In this particular sample
> network I'm guessing that there may be about an extra 7-8 ms for IPv6
> requests to Google (this is end-to-end, so the sources of latency
> could be multiple including whether or not the interconnect paths for
> IPv4 and IPv6 are the same).
> 
> Lastly, I am attaching an old slide from a measurements presentation
> we gave at the V6 World Congress in Paris in the very beginning of
> this year wherein I attempted to graphically illustrate the effect of
> different HE implementations.
> <Finicky Eyeballs.png>