Re: [tcpPrague] [tsvwg] ecn-l4s-id: Proposed Changed to Normative Classic ECN detection Text

["De Schepper, Koen (Nokia - BE/Antwerp)" <[e-mail]>]

Hi Christian,

Indeed a very important discussion we need to have, and agree on.

>> If you give up the formula, you could end up with a significantly simpler event based congestion control, in which applications slow down if they see too many CE marks, accelerate if they only see a very low marking rate, and stay put if they see an infrequent marking rate.

This is exactly what I expect most CCs will do: work event based. The formula is only to tune the events so it converges to the formula (eventually). If there is a lot of dynamism, obviously shorter RTTs can respond faster and will benefit from that. They will have memory over a shorter time interval, and on average more throughput than a longer RTT flow. But if things are stable, it is good to converge to a known rate. I don’t see any better solution, do you? If we don’t foresee such a mechanism how do you expect bandwidth to be shared over shared infrastructure? Then the whole discussion of fairness can be dismissed, and CCs should not be evaluated for this neither.

>> The marking rate is actually 0.1%, about 2 packets per RTT

Indeed, this is still 2 marks per 15ms, much better than 1 drop/mark per 1000 RTTs (15 seconds) for your example when using Reno (Cubic will also be seconds). Only if the RTT is much smaller than 15ms, it will get less marks per RTT, but in our Linux Prague we have an option to gradually take the RTT dependence into account, depending on how long the flow is running application unlimited (or by extension how long things are stable).

Also such a formula is just a reference to aim at. It should not be a continuous hard requirement. There are reasons when a “flow” can deviate from it (Less than best effort can take less, a tunnel of multiple flows will take as many times more as there are flows inside, and a real-time application might temporary take more when needed).

I’m in favor of specifying explicitly such a formula as a target for L4S CC’s to share BW in steady state situations (long flows and stable throughput). Up to now I don’t think we have such an explicit requirement in the drafts. Anyone else that would support this? At least CC control developers then have a target to aim for and a clear evaluation criterium.

Regards,
Koen.


From: Christian Huitema <[e-mail]>
Sent: Monday, November 2, 2020 1:11 AM
To: De Schepper, Koen (Nokia - BE/Antwerp) <[e-mail]>; Jonathan Morton <[e-mail]>
Cc: iccrg IRTF list <[e-mail]>; Bob Briscoe <[e-mail]>; tsvwg IETF list <[e-mail]>; TCP Prague List <[e-mail]>
Subject: Re: [tcpPrague] [tsvwg] ecn-l4s-id: Proposed Changed to Normative Classic ECN detection Text



On 11/1/2020 3:55 PM, Christian Huitema wrote:


On 11/1/2020 1:53 PM, De Schepper, Koen (Nokia - BE/Antwerp) wrote:

>>By the way, I do think that L4S makes too many hypotheses about the CC. I wish the AQM was based on intrinsic properties, such as "this flow is using more than its fair share of bandwidth"

This last quote is exactly what the L4S marking is intended to do: the smoothed marking probability over a longer time interval (about 10 RTTs) indicates the fair target rate. Assuming we agree on 15ms being the reference RTT, then the average rate can be r=2/(p*0.015)=133/p packets per second. So you can evaluate yourself if you use more or less than the fair share. How you respond to this knowledge is up to your CC.

I think that part of the complexity in L4S AQM comes from trying to achieve fair sharing of resource without actually implementing fair queuing. You end up expecting that implementations apply a formula and derive their packet sending rate from a smoothed average of the packet marking rate. It may work well in controlled environments, but I am afraid that very few implementations are ever going to do that in practice.

Congestion Control algorithms are much more likely to treat packet losses, markings, or delay increases as signals that they are sending too fast, while considering their absence or lower frequency as signal that they might be able to send faster. Implementations also are likely to take into account the recent history, for example monitoring the min RTT (LEDBAT, BBR), or monitoring the recent available bandwidth (BBR). They also typically filter signals, e.g., consider only one packet loss or ECN mark per RTT (New Reno, Cubic, etc.). This is expected, because we do see packet losses arriving in batches, and can reasonably expect ECN marks to have the same behavior. But implementations are inherently selfish, do not really trust the signals from the network, and are very unlikely to end up applying the formula that you suggest.

In any case, you have a scaling issue. Let's consider a 1.5Gbps link, with 15 ms delay and 1500 bytes packets. The nominal sending rate is 125,000 packets per second. The marking rate with your formula shall be p = 2/(r*0.015), which is about 0.0008%. Over the last 10 RTT, the connection will on average see 0.14 marks -- that is, no mark over the last 10 RTT 86% of the time. This falls well short of the requirement to provide frequent feedback!

Sorry, bug in my spreadsheet. The marking rate is actually 0.1%, about 2 packets per RTT. Still not frequent enough for my taste, but much better than 0.0008%. Nevertheless, the inverse relation between marking rate and data rate is not great.

If you give up the formula, you could end up with a significantly simpler event based congestion control, in which applications slow down if they see too many CE marks, accelerate if they only see a very low marking rate, and stay put if they see an infrequent marking rate. I can see applications doing that, resulting in a fairly stable network and low delays. Of course, you will have to do some form of fair share enforcement to catch violators, but I am sure you can come up with something.

-- Christian Huitema







_______________________________________________

tcpPrague mailing list

[e-mail]<mailto:[e-mail]>

https://www.ietf.org/mailman/listinfo/tcpprague