Re: [tsvwg] Reasons for WGLC/RFC asap

[Greg White <[e-mail]>]

Hi Pete,

Thanks for summarizing your concerns.

As others have pointed out, it is believed that the vast majority of RFC3168 implementations are fq_codel/CAKE instances, so this is an important situation to understand.

I think that it has also been pointed out that these implementations are all software upgradable, and many of them are run on systems that automatically (or readily at least) receive updates that address security vulnerabilities, fix bugs, make performance or feature enhancements, etc.

Is it possible that in these software implementations, the LSB of the ECN field can be added to the FQ flow identifier, thus allowing the FQ to segregate the L4S traffic in the tunnel from the Classic traffic?  If so, would that solve this problem?  Is that an experiment that you’d be willing to try?

It would be good also to test this with the low CE threshold enabled for queues that contain ECN==*1 as well.  These seem like two really simple changes that could propagate into the fq_codel/CAKE deployed base along the timelines that it will take for L4S experimentation to ramp up.

-Greg

From: tsvwg <[e-mail]> on behalf of Pete Heist <[e-mail]>
Date: Thursday, November 19, 2020 at 1:03 PM
To: Gorry Fairhurst <[e-mail]>
Cc: tsvwg IETF list <[e-mail]>
Subject: Re: [tsvwg] Reasons for WGLC/RFC asap

On Thu, 2020-11-19 at 16:34 +0000, Gorry Fairhurst wrote:
On 19/11/2020 16:22, Pete Heist wrote:

Hi Koen,

Rather than thinking of this as advantages and disadvantages to waiting, I see it as an engineering process. It was decided earlier this year that the L4S proposal has enough support to continue, so we're on that path now. Part of that decision, as I understood it, also recognized that there are valid safety concerns around compatibility with existing AQMs, and some solution needs to be devised.

RFC3168 bottleneck detection was added to TCP Prague, which appears to be difficult to do reliably when there is jitter or cross-flow traffic, and it has since been disabled in the reference implementation. The l4s-ops draft was started, but isn't complete yet and may need WG adoption as part of a LC. We can then decide how effective the proposed mitigations are against the risks and prevalence.

To start a WGLC now would circumvent that earlier recognition that a safety case needs to be made. Meanwhile, since testing showed that tunnels through RFC3168 FQ AQMs are a straightforward path to unsafe flow interaction, along with other issues relative to the goals, it doesn't seem like the engineering process is done just yet.


By the way, I liked your data - and it helped me a lot to look at this, thanks very much for doing this.
I'm glad, as I think we're at our best when we're doing engineering and producing data. I wish it were easier to do!

It would help me if you clarify what you mean by  "unsafe" - to me "safety" relates to traffic unresponsive to drop, as in CBR traffic, etc. I've not understood how CE-marked traffic can erode safety, but maybe I missed something?
Sure, so the existing RFC3168 CE signal in use on the Internet today indicates an MD (multiplicative decrease), whereas the redefined CE signal in L4S indicates an AD (additive decrease). Two congestion controls responding to CE in a different way, or one that responds to CE with an AD and one that responds only to drop (i.e. all standard congestion controls that advertise Not-ECT), will not interact safely in the same RFC3168 signaling queue. We're probably on the same page here already, but I'll refer to section 5 of RFC8257.

That is one of the reasons why ECT(1) is used in L4S to place L4S flows in the L queue- to keep them separate from conventional flows in the C queue. As long as flows have advertised their capability correctly, that works.

However, existing RFC3168 queues do not have knowledge of L4S, therefore will not know that ECT(1) means that traffic needs to be segregated and signaled in a different way. They will signal a Prague flow, which sets ECT(1), with CE, expecting the flow to respond with an MD, rather than AD. Meanwhile they'll signal an RFC3168 or non-ECN flow with either CE or drop, and in either case the flow will respond with an MD, causing conventional flows to yield to Prague flows to varying degrees depending on the AQM in use.

Here's an example of CUBIC and Prague when they end up in the same fq_codel queue:
http://sce.dnsmgr.net/results/l4s-2020-11-11T120000-final/l4s-s6-rfc3168-1q/l4s-s6-rfc3168-1q/l4s-s6-rfc3168-1q-ns-prague-vs-cubic-fq_codel-50Mbit-20ms_tcp_delivery_with_rtt.svg

Here's a more extreme example of Reno and Prague sharing a single PIE queue with ECN enabled (less common):
http://sce.dnsmgr.net/results/l4s-2020-11-11T120000-final/l4s-s6-rfc3168-1q/l4s-s6-rfc3168-1q/l4s-s6-rfc3168-1q-ns-prague-vs-reno-pie-50Mbit-20ms_tcp_delivery_with_rtt.svg

In the example with PIE, Reno appears to be driven at or close to minimum cwnd. In the fq_codel example, the steady state throughput of Prague:CUBIC is around 19:1. We've seen a range in the Codel case from around 12:1 to 20:1. In my opinion, we could use the word "unsafe" here in both cases.

I'm not sure why "tunnels have crept in here. There have always been side-effects with classification (and hence scheduling), but I don't see new issues relating to "tunnels" with ECN.
Tunnels are relevant because they provide an easy practical path to the unsafe flow interaction described above. The widely used fq_codel qdisc has ECN enabled by default. Fortunately, because it has flow-fair queueing, Prague flows and conventional flows are usually placed in a separate queue (hash collisions aside), causing Prague to only affect itself with additional delay (TCP RTT). However, a tunnel's encapsulated packets all share the same fq_codel queue because they all have the same 5-tuple, so there is unsafe interaction between the tunnel's flows. Here we use Wireguard through fq_codel:

http://sce.dnsmgr.net/results/l4s-2020-11-11T120000-final/l4s-s5-tunnel/l4s-s5-tunnel-phys-wireguard-prague-vs-cubic-fq_codel-50Mbit-20ms_tcp_delivery_with_rtt.svg

I'll leave it to the WG to come up with examples of what types of tunnels and traffic scenarios could lead to this, but one example is a user who has a privacy VPN on their PC, and fq_codel on their home gateway. Let's say one flow connects to an L4S capable server, and another flow to a non-L4S, conventional server. The L4S flow will dominate the non-L4S one (whether it's ECN capable or not), probably causing some level of poor service, perhaps for a video stream, download, or whatever.

I'm not commenting on when the Chairs think a WGLC will provide useful information, we'll say that in due course.
Ok, I trust that we'll engage enough disinterested people into congestion control who will add their input.

Thanks Gorry for looking this over. :)

Best wishes,

Gorry

Regards,
Pete

On Wed, 2020-11-18 at 10:31 +0000, De Schepper, Koen (Nokia - BE/Antwerp) wrote:

Hi all,

To continue on the discussions in the meeting, a recap and some extra thoughts. Did I miss some arguments?

Benefits to go for WGLC/RFC asap:

·         There is NOW a big need for solutions that can support Low Latency for new Interactive applications

·         The big L4S benefits were a good reason to justify the extra network effort to finally implement ECN in general and AQMs in network equipment

·         Timing is optimal now: implementations in NW equipment are coming and deployment can start now

·         Deployment of L4S support will include deployment of Classic ECN too! So even for the skeptics among us, that consider that the experiment can fail due to CCs not performing to expectations, we will fall back to having Classic ECN support

·         Current drafts are about the network part, and are ready and stable for a very long time now.

·         Only dependency to CCs in the drafts are the mandatory Prague requirements (only required input/review from future CC developers: are they feasible for you)

·         We have a good baseline for a CC (upstreaming to Linux is blocked by the non-RFC status)

·         Larger scale (outside the lab) experiments are blocked by non-RFCs status

·         It will create the required traction within the CC community to come up with improvements (if needed at all for the applications that would benefit from it; applications that don’t benefit from it yet, can/will not use it)

·         NW operators have benefits now (classic ECN and good AQMs) and in the future can offer their customers better Low Latency experience for the popular interactive applications

·         When more L4S CCs are developed, the real independent evaluation of those can start

Disadvantages to wait for WGLC/RFC:

·         We’ll gets stuck in an analysis paralysis (aren’t we already?)

·         Trust in L4S will vanish

·         No signs that we can expect more traction in CC development; trust and expectations of continuous delays will not attract people working on it, as there will be plenty of time before deployments are materializing

·         Product development of L4S will stall and die due to uncertainty on if L4S will finally materialize

·         Product development of Classic ECN will stall and die due to uncertainty on how L4S will finally materialize

What are the advantages to wait? Do they overcome these disadvantages?

Regards,
Koen.




-->