Re: [v6ops] PMTUD issue discussion

[joel jaeggli <[e-mail]>]

On 8/30/14 11:36 PM, Ray Hunter wrote:
>> Matthew Petach <mailto:[e-mail]>
...

> Not necessarily.
> 
> What might help is a realistic and early estimate of the minimum MTU 
> we're likely going to encounter on the path,. even if this is just
> used as a hint for the PLMTUD learning process that is communicated
> from the local router to local hosts.

The case where we would see this occur typically involved encapsulation,
It would be the tunnel ingress on the way to the client  that sends the
PTB message.  The client is practically speaking unaware that the MTU is
constrained (even if the immediate upstream device is aware (because it
terminates the tunnel). in that sense it may be a lot simpler to inform
the client that the mtu is constrained e.g. by setting it in the RA even
though the lan segment can in fact support something larger (e.g. rfc
4861s observation about heterogeneous networks).

You as the client could be several hops from your end of the tunnel I
suppose.

> MTU is configured on routers. Routers determine the path (based on 
> routes). Routes are based on prefixes. So there is certainly a
> relationship between MTU and prefixes.
> 
> Your concern seems to be how good that correlation is.
> 
> IMVHO I believe we could generate a single minimum MTU per path
> (based on prefixes) that is a pretty good first approximation (whilst
> avoiding blowing up the routing table).

So In my specific case I actually do have a full table on servers so
that I can do outbound path selection, per prefix/as statistics for the
application and so on, that's maybe  an unnecessary set of information
for most "servers" e.g. devices that receive incoming connections.

> In EIGRP, a lot of AS'es don't even use summary routes. So in this
> case there's a very good correlation between MTU, path, and prefix,
> at least for traffic within the AS. So I could see this working quite
> well within an individual AS, like a corporate network.
> 
> In BGP4, the path is a set of AS'es to be traversed.
> 
> The minimum MTU of an individual AS would be the minimum MTU of all 
> individual paths encountered when traversing that AS. e.g. the
> minimum of all MTUs learned in the IGP, or a static manually 
> configured number in the AS border router.
> 
> The minimum MTU of a BGP path would be the minimum of the individual
> AS MTUs.
> 
> I can see this being potentially problematic, but is it meaningfully 
> better than a single global constant (based on a minimum of the
> whole Internet)?
> 
> Or is a search for the path MTU for every host for every session
> already a good enough solution?

connections are frequently pretty short-lived I personally like to
minimize the number of times that we need to engage in negotiation when
setting up a tcp connection. fortunately as I noted this case is
actually kind of infrequent and hopefully even less common as the
network matures.