Re: [tcpm] Possible error in accurate-ecn

[Bob Briscoe <[e-mail]>]

Mirja, Richard, Yoshi,

I just posted a new rev of AccECN for a number of other reasons, and 
realized we hadn't converged on an answer to the tricky problem in this 
old thread from Nov'20. For now, the text says the following (which I've 
justified below, but I don't think it's the solution yet), and I've 
asked Richard & Mirja if we (as co-authors) can act as a little design 
team to thrash out this problem in the next few days.

3.2.2.5.1.  Data Receiver Safety Procedures

    An AccECN Data Receiver:

    o  SHOULD immediately send an ACK whenever a data packet marked CE
       arrives after the previous packet was not CE.

    o  MUST immediately send an ACK once 'n' CE marks have arrived since
       the previous ACK, where 'n' SHOULD be 2 and MUST be in the range 2
       to 6 inclusive.


Justification for first bullet: The aim is to trigger an ACK on a 
transition from non-CE to CE, but only when the latest packet is a data 
packet. Cases where the previous packet was a CE-marked pure ACK and 
therefore not ACK'd ought to be picked up by the 2nd bullet, not the 
first. Similarly, if there's CE-marked packets (data or pure ACKs) that 
haven't been ACKd yet, and the next data packet is not CE-marked, I 
think the first bullet is still correct, and the second rule will pick 
up cases with enough CE marks.

The second bullet is more tricky. I've added Richard's suggesting of 
flooring 'n' at 2. But made no other changes. That's deliberate, 'cos I 
think it's sufficient to allow ACK CC to be added, but not to preclude 
it. And I think the danger of being interpreted as DupACKs shouldn't 
apply if there's no outstanding data. I know I'm wrong in those odd 
cases I mentioned when new data might cross the ACKs in the other 
direction. But I don't think that's going to cause too much harm.

However, I'm probably very wrong. Somehow I feel we haven't got a good 
model to check all the possibilities.


Bob



On 17/11/2020 08:48, Scheffenegger, Richard wrote:
>
> See my suggestion; The critical requirement for the 2nd bullet point is,
> to have n > 1 for pure, CE-marked ACKs. The exponential decay, after how
> many ACKs triggered by CE-marked ACKs these ACK-for-ACK stop, strongly
> depends on "n", so the recommendation should be to use a higher value of
> n in the case where you only observe a stream of pure ACKs (with or
> without marks; only those marked would account against "n" though).
>
> In the case of data packets mixed in, you don't want to delay
> excessively, thus "n" may be choosen lower...
>
> Richard
>
>
>
> Am 10.11.2020 um 17:38 schrieb Bob Briscoe:
>>
>> Moving on to the second bullet...
>>>> The second bullet doesn't consider the possibility that the 'n'th CE
>>>> mark might arrive on a pure ACK. Then, the wording as it stands says
>>>> the Data Receiver MUST immediately ACK a pure ACK. I know TCP never
>>>> ACKs a pure ACK, but I'm not actually sure it does any harm to do so
>>>> in this case (it cannot cause an infinite loop of ACKs). However,
>>>> given it would be unorthodox, we maybe ought to rule it out by
>>>> rewording anyway?
>>> [MK] I think we also can leave this as it is because if that ’n’
>>> packet is a pure ACK that still means that you have unacked data
>>> packets with CE marks and you should trigger an ACK for those packet
>>> now (rather than waiting for the delayed ACK timer to expire). However
>>> this case is less important. Also we should probably make sure that
>>> this doesn’t apply if there are only pure ACKs with CE marks, maybe by
>>> add “if unacknowledged data are outstanding” or something.
>>
>> [BB] By the end of your last para you start to realize that, if the n'th
>> CE mark arrives on pure ACK, it doesn't say anything about whether the
>> previous (n-1) CE marks were on data or pure ACKs.
>>
>> Consider another common scenario; a server delivering chunks of video
>> over a high BDP path (e.g. 500 packets per RTT), and going idle between
>> times (perhaps in response to an "exceeded high watermark" data packet
>> from the server). Now consider the possibility that the reverse path
>> bottleneck builds a queue during each chunk, and starts CE marking
>> towards the end of each chunk. Then, after the server stops sending a
>> chunk, it continues to see a couple of hundred CE-marked pure ACKs
>> arriving.
>>
>> However, by your proposed rule, it doesn't send any feedback, 'cos it
>> has no data to acknowledge. So the client cannot regulate its ACK
>> stream. That's just tough, 'cos the client will stop sending more pure
>> ACKs after it has received the last data, which is before any later
>> feedback from the server can reach it.
>>
>> However, if the client later sends a data packet that the server can
>> acknowledge (e.g. "a low watermark" message), the server's ACE counter
>> will have wrapped dozens of times, and the client won't be able to work
>> out how many. That doesn't matter if a few round trips have elapsed -
>> the congestion info will have become stale. But it does matter if it's
>> fairly soon.
>>
>> Instead, if we keep the text of the second bullet... if the server's 'n'
>> was 2, it would send a pure ACK for every other pure ACK it received.
>> But, what if these pure ACKs from the server were also CE-marked? Then
>> the client would feed back pure ACKs using its 'n'. This regression
>> would eventually die out, but I don't like it.
>>
>> This conversation about the second bullet presumes AccECN is being used
>> for ACK congestion control. I'm not comfortable with writing anything
>> specific about ACK congestion control into a draft on the standards
>> track. But I don't want to preclude AccECN being used for ACK CC.
>>
>>
>> I believe the second bullet needs something changed to limit the ACKing
>> of pure ACKs. However, before we put effort into wordsmithing, I'd like
>> to hear whether you agree with the general idea of not defining what to
>> do for ACK CC, but not precluding it.
>>
>> Cheers
>>
>>
>>
>>
>> Bob
>>
>>>
>>> Mirja
>>>
>>>
>>>
>>>> Thoughts anyone?
>>>>
>>>>
>>>> Bob
>>>>
>>>> -- 
>>>> ________________________________________________________________
>>>> Bob Briscoe
>>>> http://bobbriscoe.net/
>>> _______________________________________________
>>> tcpm mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/tcpm
>>
>
> _______________________________________________
> tcpm mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tcpm

-- 
________________________________________________________________
Bob Briscoe                               http://bobbriscoe.net/