Re: [multipathtcp] Draft MPTCP Charter

[marcelo bagnulo braun <[e-mail]>]

Ford, Alan escribió:
> Marcelo,
>
>   
>>>> Not really.
>>>> actually this is what you need for the two ended version.
>>>> the one ended version assumes PI addressing that can be used in
>>>>         
> both ISPs.
>   
>>>> The two ended version assumes that if a node has two PA addresses
>>>>         
> from
>   
>>>> two different ISPs, the a packet with the source address of ISPA
>>>>         
> will be
>   
>>>> routed to ISPA or that if this is not the case ISPB will propagate
>>>>         
> it
>   
>>>> (i.e. what you describe above)
>>>>
>>>>         
>>> That is absolutely not the case.
>>>
>>>       
>> really?
>>     
>
> Yes.
>   

well, no
>   
>>> In the draft, we describe the multiple TCP subflows as having
>>>       
> separate
>   
>> addressing. The source address on the packet is always valid for the
>>     
> link in
>   
>> use. If a packet is to be retransmitted on a different subflow it uses
>>     
> that
>   
>> subflow's addressing instead.
>>
>>     
>>>       
>> so, you are assuming that the two ended version of MPTCP only applies
>>     
> to
>   
>> hosts with multiple interfaces and that each interface has a different
>> IP address. this is overly restrcitive imho.
>>     
>
> Well we assume the hosts are multi-addressed, not necessarily with
> multiple interfaces. 
good, that is what i was assuming

> The draft title is "TCP Extensions for Multipath
> Operation with Multiple Addresses" which I thought made it reasonably
> clear!
>
>   
>> You should be able to apply it to a host located in a multihomed site
>> and that the multihomed site is obtaining multiple PA blocks one from
>> each provider it connects to, or is this out of the scope?
>>     
>
> If the hosts were still multi-addressed that would work fine.
>
>   

not really

please see http://www.shim6.org/draft-ietf-shim6-ingress-filtering-00.txt
>> If you intend to support this scenario, then the problem that i
>>     
> describe
>   
>> above certainly occurs.
>>     
>
> Well I guess that depends on the behaviour of the egress router. If
>   
> there were separate egress routers for each ISP, this wouldn't be a
> problem. If it was the same box, so long as it manages routing
> appropriate (and it probably would so as not to hit the upstream ISP's
> RPF), it would still work since the router would send the right packets
> to the right ISP.
>   

the problem is that this doesn't happen,
What you are describing is described in the draft that i pointed above 
and relies on using source address dependent routing in the end site, 
which is by far a not widely implemented technique.

But the bottom line, that i was pointing out that started all this 
discussion, is that the one ended approach does _not_ have this problem 
cause the host has a single PI address. It does happen in some 
configurations in the two ended approch. Do we agree?

Regards, marcelo

> Regards,
> Alan
>
>
>