[Battlemesh] Session about multi-homed IPv6 mesh networks and auto-configuration

[Ferry Huberts]

On 19/04/16 18:48, Dave Taht wrote:
> On Tue, Apr 19, 2016 at 9:40 AM, Ferry Huberts <mailings at hupie.com> wrote:
>>
>>
>> On 19/04/16 18:23, Juliusz Chroboczek wrote:
>>>>
>>>> We need metrics in the homenet protocols to be able to handle this.
>>>
>>>
>>> Homenet uses Babel for announcing reachability, which has metrics.
>>
>>
>> Yes, but AFAIK there are no provisions for announcing exit link bandwidth.
>>
>> We really need something like that to be able to choose the best gateway to
>> use. This becomes especially relevant in mesh scenarios where nodes might be
>> multiple hops away.
>>
>> As an example: a mesh node is 1 hop away from an exit link with very little
>> bandwidth, and several hops away from an exit link with a very high
>> bandwidth. Depending on conditions it might be advantageous to pick either
>> one or the other. This is only possible when we know BOTH the exit link
>> bandwidth and the mesh metrics.
>
> I prefer to think about it as a congestion problem. If that local link
> is entirely unused, and that gigE link saturated, which would you
> chose?

Exactly.

Nobody said those bandwidth are static ;-)
Multi-smart-gateway is actually designed to handle this and we use it 
already this way.

This is also somewhat alleviated by MP-TCP.
But MP-TCP is not very close as a mainstream tech.


>
>> This becomes somewhat less relevant when using MP-TCP. Not very much however
>> when there are many exit links since you generally don't want to use many
>> many exit links in a MP-TCP session, especially on a mesh.
>
> Source specific routing pushes the choice of exit link to the
> application, which I like a lot. http://arxiv.org/pdf/1502.02402.pdf

Me too, but the application needs information to decide upon.
Right now it doesn't have the right information IMHO.

> for mosh was quite lovely (did it ever make mainline?), and mp-tcp is
> not required - IF we could push source address selection into the
> application for regular tcp (or quic), in a browser, something much
> like happy eyeballs would suffice.
>
>>
>>
>>>
>>> Where Homenet is different from what you're used to is that it announces
>>> all available gateways using source-specific routes, and hosts are
>>> supposed to choose the gateway they use (by picking a given source
>>> address).  See
>>>
>>>     http://arxiv.org/pdf/1403.0445.pdf
>>>
>>> Unless they snoop on Babel traffic, however, hosts have no information
>>> about the metrics associated to individual prefixes.  Snooping on Babel
>>> traffic is not difficult, but it is not defined by the Homenet
>>> architecture,
>>> since Homenet requires no host changes.
>>>
>>> -- Juliusz
>>> _______________________________________________
>>> Battlemesh mailing list
>>> Battlemesh at ml.ninux.org
>>> http://ml.ninux.org/mailman/listinfo/battlemesh
>>>
>>
>> --
>> Ferry Huberts
>>
>> _______________________________________________
>> Battlemesh mailing list
>> Battlemesh at ml.ninux.org
>> http://ml.ninux.org/mailman/listinfo/battlemesh
>
>
>

-- 
Ferry Huberts