Re: [tsvwg] L4S DSCP (was: L4S drafts: Next Steps)

[Pete Heist <[e-mail]>]

I'll just add to the sentiment that I think the use of DSCP is still
worthy of consideration. Beyond the use of a single DSCP on all
traffic, there may be other alternatives that address at least some of
the concerns in B.4.

Pete

On Wed, 2021-03-24 at 12:17 +0200, Jonathan Morton wrote:
> > On 24 Mar, 2021, at 10:28 am, Gorry Fairhurst
> > <[e-mail]> wrote:
> > 
> > This really  is not a new topic, and the WG should only be
> > revisiting this if there is a new requirement - the design to allow
> > the traffic to be DSCP-marked was not taken without thought and 
> > visibility.
> 
> I agree that it's not a new topic.  However, it has renewed relevance
> with the introduction of the l4s-ops draft, since L4S has failed to
> adequately address the open issues which prevent its being deployed
> as an Internet-wide experiment (which is the predicate for many of
> the arguments against DSCP in B.4).
> 
> Since David Black brought up RFC-4774 (a remarkably prescient
> document), I will note here that it divides alternate-ECN proposals
> into three possible categories:
> 
> Option 1: cannot be deployed on the Internet.  This clearly describes
> DCTCP, for example.
> 
> Option 2: can reliably detect whether it is receiving RFC-3168 or
> "alternate" signals and switch modes gracefully.
> 
> Option 3: is inherently compatible with RFC-3168 traffic and
> infrastructure.
> 
> There is no Option 4 contemplated; certainly none where harm to or
> required action by innocent bystanders is accepted.
> 
> I appreciate that L4S intended to improve DCTCP sufficiently to merit
> an Option 2 classification.  However, it has clearly failed to do so
> to date, despite man-years of effort from its proponents and many
> hours of WG meeting time expounding its presumed benefits (while
> ignoring warnings of its downsides).  This leaves L4S as an Option 1
> proposal, though with less restricted applicability than DCTCP since
> it does at least carry an identifier; it can only safely be deployed
> on networks that have been prepared for it.
> 
> This means a containment mechanism is required, and DSCP is perfect
> for the job.  The only alternative I see is to write l4s-ops - or
> include language in the other drafts - such that it requires
> deployment only in similarly isolated circumstances as DCTCP.  The
> latter option would be unpalatable, I'm sure, but people really have
> painted themselves into a corner on this issue.
> 
>  - Jonathan Morton