Sheng, Pls see [BB]... On 26/08/2022 11:03, Sheng Jiang via Datatracker wrote:
Reviewer: Sheng Jiang Review result: Has Issues Reviewer: Sheng Jiang Review result: Has Issues Document: draft-ietf-tsvwg-aqm-dualq-coupled-24 Review Date: 2022-08-26 I have reviewed this document as part of the Operational directorate's ongoing effort to review all IETF documents being processed by the IESG. These comments were written with the intent of improving the operational aspects of the IETF drafts. Comments that are not addressed in last call may be included in AD reviews during the IESG review. Document editors and WG chairs should treat these comments just like any other last call comments. This experimental document defines a framework for coupling the AQM algorithms in two queues intended for flows with different responses to congestion. It is the network part of the L4S architecture that enables both very low queuing latency and high throughput at the same time. I have concerns on the operational description in this document. This document claims its mechanims can be deployed in the current network. But it lacks the description on how and why it can works back-compatibly. I heard there were intensive debate regarding to the usage of ECT (I did not participate it myself), but the document has not introduced the reason or background why it is allowed to use ECT from the urgement.
[BB] I'm afraid this comment is directed at the wrong draft, because it confuses two things: * This document (DualQ) is about where a bottleneck buffer /has/ been updated with L4S support. The DualQ draft is entirely about a backwards compatibility mechanism. * The intensive debates were about possible cases where a bottleneck buffer had previously supported Classic RFC3168 ECN, and it had /not/ been updated for L4S.
The remedies for the second case are in the requirements for L4S sending hosts, which is where that debate was directed ( draft-ietf-tsvwg-ecn-l4s-id ). The WG has now moved past that debate, onto the next stage, which is to measure the size and severity of this issue, experiment with the required backward compatibility mechanisms in senders and to record all this in the l4sops draft.
The main question was how prevalent cases might be where L4S and non-L4S flows could compete in a queue that supports Classic RFC3168 ECN, versus whether they would nearly always be separated by a per-flow scheduler, and how serious and how sustained the problem would be if they did mix sometimes. That's nothing to do with this DualQ draft.
Another minor comments, there is another reason that can cause the result mentioned in 4.2: if irresponsible terminals labels classic ect(0) flows into l4s ect(1), it would overload l4s queue.
[BB] That's not the case with this DualQ system. The two queues feed into the same capacity, and either queue can use it all. So there's no more overload if traffic is switched from one to the other, even if source(s) are unresponsive. And the system is designed so that the priority scheduler gives no advantage to overload with ECT1. Because if the total load into both queues is greater than the capacity, the overload mechanism kicks in. It couples the two queues together as one, and applies an equal proportion of AQM drop to both. Here's the overload experiments showing that, which were explained in the WG: https://l4steam.github.io/overload-results/
Thank you for taking the time to write up your review. Bob
Regards, Sheng
-- ________________________________________________________________ Bob Briscoe http://bobbriscoe.net/ -- last-call mailing list last-call@xxxxxxxx https://www.ietf.org/mailman/listinfo/last-call
