Re: [multipathtcp] Multipath TCP Address advertisement 2/5 - Reliability

[Yoshifumi Nishida <[e-mail]>]

Hi Alan,

On Thu, Aug 4, 2016 at 12:57 AM, Alan Ford <[e-mail]> wrote:

> Hi Yoshi,
>
> Inline…
>
> On 4 Aug 2016, at 07:36, Yoshifumi Nishida <[e-mail]> wrote:
>
> Hi Fabien,
>
> On Wed, Aug 3, 2016 at 8:47 AM, Fabien Duchêne <fabien.duchene@
> uclouvain.be> wrote:
>
>> Hello,
>>
>> As agreed in Berlin during IETF96, I'm sending a series of emails to
>> discuss the different contributions proposed
>> inhttps://datatracker.ietf.org/doc/draft-duchene-mptcp-add-addr/
>>
>> This is the part 2/5 : reliability.
>>
>> In RFC6824, ADD_ADDR options can be attached to segments carrying data
>> or pure acknowledgements.
>> In practice, notably given the length of ADD_ADDR with IPv6 addresses
>> and the HMAC, it is very likely that they will be often sent as pure
>> acknowledgements.
>> This implies that ADD_ADDR are sent unreliably, which could be
>> problematic when the ADD_ADDR is required to allow the establishment of
>> additional subflows, as in load balancing scenarios.
>> We propose to rely on the "E" (Echo) flag in the ADD_ADDR option.
>> This echo flag is used to acknolwedge a received ADD_ADDR by echoing it.
>> If the acknowledgement is not received, the ADD_ADDR option will be
>> retransmitted up to N times.
>>
>
> I have several comments on this. Please let me know if I miss something.
>
> 1： Do we really need ADD_ADDR reliability in all cases? An end node might
> want to send ADD_ADDR for 'just in case' rather than for "I want you to use
> this"
> BTW, I'm not sure if the draft tries to replace the current ADD_ADDR or to
> propose an additional ADD_ADDR option.
>
>
> The argument is that by making it reliable, you only need to send it once
> to know it’s been received, even if the far end does not use it.
>
2: The sender cannot be sure whether the info in ADD_ADDR will be used or
> not. It depends on the peer's decision.
> I might prefer re-transmitting ADD_ADDR up to N times when the sender
> doesn't receive MP_JOIN to the address for a certain amount of time.
>
> As above, by making it reliable you only need to send it once to know it’s
> been received, even if the far end chooses not to use it.
>

Hmm. if a sender wants the peer to use specific address (such as a load
balancer case), I can understand that we want to make it reliable. But, if
the sender doesn't care, I'm not very sure the reason to make it reliable
or send ADD_ADDR frequently.

I personally think this is a trade-off point. We are trying to introduce
some complexities in the protocol to send TCP options reliably.  I might
want to see some benefits to compensate it.


>
> I think I’m sold on this idea now, given the additional recent comments. I
> support this.
>
> 3: "the receiving host MUST return the exact option.." Although it is
> MUST, can we always do this? There might be a situation where there is no
> option space or no segment to send.
>
>
> I would assume the same rules as ADD_ADDR are followed, so it’s sent on a
> duplicate ACK in that case.
>

Hmm. If we apply the same rules to this, it may need to send a dupack for a
dupack and may be repeated N times. I still need to think about it, but it
doesn't look pretty..

Thanks,
--
Yoshi