Re: [tsvwg] plan for L4S issue #29

[Pete Heist <[e-mail]>]

On Thu, 2020-10-01 at 13:29 +0100, Gorry Fairhurst wrote:
> On 01/10/2020 13:13, Pete Heist wrote:
> > On Tue, 2020-09-29 at 22:48 +0000, Greg White wrote:
> > > In my opinion, Issue 29 can be closed. RFC3168 detection should
> > > continue to evolve and should be referenced in the Operational
> > > Guidance draft, but fallback should not be required in the
> > > experimental protocol drafts for the reasons outlined in Issue
> > > 29,
> > > and on the mailing list.
> > In my opinion, the issues with bottleneck detection and fallback
> > should
> > be fixed, since this was the proposed way to provide safety for the
> > alternate meaning of CE being introduced.
> > 
> > The question of whether or not to proceed with these issues still
> > open
> > sounds like a separate one, but it seems like closing issues
> > without
> > fixing them would make it hard for someone who wanted to review and
> > understand what issues are still open.
> > 
> > Pete
> > 
> To be clear, is there a "safety" concern with respect to congestion 
> collapse from overload, or is it an issue concerning relative
> fairness 
> between flows?

False detection (this issue #29) is by itself less serious directly
than missed detection, because the failure mode is safe interaction
with RFC 3168 flows. However, false detection events become an
incentive to disable detection, so this issue is still part of a chain,
unless detection and fallback is required.

There's a plot over in issue #16 of one example of what happens when
RFC 3168 flows and L4S flows wind up in the same Codel queue (however
that happens) either without detection and fallback enabled, or if a
missed detection occurs:

https://trac.ietf.org/trac/tsvwg/ticket/16

Steady state throughput in that case appears to be about 44Mbps for the
L4S flow and 4Mbps for the RFC 3168 flows. Someone could do more
testing or predictions on how things change under different network
conditions, but a ratio of at least 16:1 is probably possible somehow,
since that was shown in some earlier slides.

In my opinion, approximate fairness is ideal, a result of 2-3x under
some conditions may still be tolerable, and anything beyond that isn't
acceptable. I remember one definition of "starvation" as "you know it
when you see it". Personally, a 10:1 throughput ratio already meets
that definition for me, so qualifies as "unsafe".

Pete