I think this is the kernel of something useful, a few comments in-line.

> On Mar 13, 2013, at 2:28 PM, <[email protected]>
>  wrote:
>
>> Fred,
>>
>> Comments below:
>>
>> Section 2, pt 2
>> "Deployed AQM SHOULD use ECN as well as loss, and set thresholds
>> to mark traffic earlier than it is lost."
>> - This is not clear, I agree SHOULD use ECN for ECT traffic, of course.
>> - I'm not sure about threshold question that sets ECN drop before ECN
>> loss
>
> As I just said to Mikael, I don't think I worded that one well. I'm
> envisaging two thresholds, testing two stress levels. It a lower one, one
> marks ECN-capable traffic and drops non-ECN-capable traffic; at the higher
> one, one drops from all traffic. What "threshold" means in a given
> algorithm is algorithm-dependent, however.
>
That's what I thought you intended.

>> - I like the idea for various reasons (I'm not expanding that here), but
>> this isn't what I understand as the current recommended TCP ECN reaction
>> -
>> which reacts to CE in the same way as loss?
>
> Well, 3168 says that TCP should respond in the same way it responds to
> loss. By that, what they mean is that it reduces cwnd, and if in
> slow-start, exit slow-start. The other thing it does to respond to loss is
> to retransmit a message. There is obviously no need to retransmit, but it
> should indeed play with cwnd in some way.
>
Agree

> How it responds will be specific to the congestion control algorithm,
> though. With newreno, it might set cwnd to 1 or to the initial value, or
> perhaps halve it.
>
1/2 would be ok

> With CUBIC, IIRC, it reduces it by a smaller fraction.
> With CalTech FAST, which is a delay-based model and tunes toward the knee,
> I could imagine that instead of literally reducing cwnd (if you're at the
> knee, why would you do that?) you might recalculate the current mean RTT.
>
>> We need to be careful that we don't suggest not using ECN can gain
>> advantage.
>
> Did I say that?
>
We need to be careful if that the incentive is correct - but you didn't
have space to explain the algorithm.

>> Section 2 pt 3
>> - Again I agree, but not sure we can say this as a BCP requirement? I
>> think we should think about how best to present this.
>
> 2.3.  AQM algorithms deployed SHOULD NOT require operational tuning
>
> There are algorithms that don't require tuning. I'm suggesting we use one.
>
If these really are shown to work on all scenarios...  Then that could be
ok, but I am not sure yet alg work in all places.

>> Section 2 pt 4
>> - Agree and we also now have tunnel technologies considering ECN
>> support,
>> so also these?
>
> 2.4.  AQM algorithms deployed SHOULD be effective on all common
>       Internet traffic
>
> We want to follow the rules for remarking traffic; I believe that's RFC
> 2983.
>
>> Section 2 pt 5
>> - Nice, but not not possible - so TCP without ECN  *IS* going to cause
>> loss and delay if it shares the same congested queue. The idea of
>> defining
>> guidance on what to expect here is also good, and maybe a significant
>> step
>> to getting a better understanding.
>
>      2.5.  TCP and SCTP congestion control algorithms SHOULD
>            maximize their use of available bandwidth without
>            incurring loss or undue round trip delay
>
> I'm not quite as pessimistic about the possibility here. We actually do
> have congestion control algorithms that tune to a point approximating the
> knee, and at least one that will co-exist with a loss-based model
> effectively (CAIA CDG). Note that I'm not dying to get to the knee,
> although that would be nice; the point is to not hit the cliff. Once the
> window advances to/beyond the knee, we are using the available bandwidth
> in its entirety; increasing cwnd after that point merely increases delay
> and the probability of loss, not throughput rate.
>
>> I note that RFC 2309 does recommend RED but importantly it did not
>> motivate it in the way that now makes AQM an imperative. It also largely
>> pre-dated ECN and certainly the experience in ECN implementation.
>
> Hmm. OK, then maybe we do need to re-do 2309. That was my first instinct
> when Wesley and I first chatted about this in mail, and I thought I would
> try simply building on it. I'll take a gander at that.
>
>> Gorry
>>
>>> Folks. I posted the email I sent yesterday as a draft, for discussion.
>>> I
>>> welcome comments, and if substantive comments are made, suggested text.
>>>
>>>
>>> On Mar 13, 2013, at 12:48 PM, <[email protected]>
>>> wrote:
>>>
>>>>
>>>> A new version of I-D, draft-baker-tsvwg-aqm-recommendation-00.txt
>>>> has been successfully submitted by Fred Baker and posted to the
>>>> IETF repository.
>>>>
>>>> Filename:   draft-baker-tsvwg-aqm-recommendation
>>>> Revision:   00
>>>> Title:              IETF Recommendations Regarding Active Queue Management
>>>> Creation date:      2013-03-13
>>>> Group:              Individual Submission
>>>> Number of pages: 7
>>>> URL:
>>>> http://www.ietf.org/internet-drafts/draft-baker-tsvwg-aqm-recommendation-00.txt
>>>> Status:
>>>> http://datatracker.ietf.org/doc/draft-baker-tsvwg-aqm-recommendation
>>>> Htmlized:
>>>> http://tools.ietf.org/html/draft-baker-tsvwg-aqm-recommendation-00
>>>>
>>>>
>>>> Abstract:
>>>>  Fifteen years after the IAB issued its recommendations regarding
>>>>  congestion control in RFC 2309, a major issue in the community is the
>>>>  issue that RFC addresses: Buffer bloat.  It may be time to update the
>>>>  recommendation.
>>>>
>>>>
>>>>
>>>>
>>>> The IETF Secretariat
>>>>
>>>
>>
>>
>




Reply via email to