Re: [tsvwg] start of WGLC on L4S drafts

[Luca Muscariello <[e-mail]>]

Hi Ruediger



On Wed, Sep 29, 2021 at 9:25 AM <[e-mail]> wrote:

> Hi Luca,
>
>
>
> I think to have dealt with VoQ features like “Approximate Fair Queueing”,
> but don’t favor them within the backbone. I think it is important to
> consider the definition of “flow” at which these systems reach the scale
> expected by the company operating them, and the presentation you’ve linked
> doesn’t define “flow” (it mentions a configurable buffer threshold to
> classify a “flow” as elephant). To offer a different sight: the operator
> wants to serve n egress interfaces and needs an appropriate number of
> queues and related counters to reliably operate a router with the desired
> load range over a long period. That impacts the number of flows, I think
> and that will impact how to define “flow”.
>

I agree with you. I was not proposing to use that ASIC in a backbone router
and indeed
Cisco does not use it there. I was just bringing that up to say that Cisco
has proved
for quite a long time that we all have reached the technology level to have
flow aware processing in silicon today. For any line rate.

If a backbone link is creating queuing, that's an issue that should be
addressed with
traffic engineering, proactively in principle, or reactively in the worst
case.
If that issue happens, for as little time as possible, relevant classes of
applications should be protected. But I also think diffserv serves well for
that need with
very predictable performance.

Best
Luca




> Dynamic Packet Prioritisation of low volume flows isn’t offered with the
> routers I’m working with, at least I’m not aware of ways to configure it.
> In general, if that’s a “two queue solution” and if a reasonable
> classification and a separate AQM can be easily configured, it might indeed
> be useful.
>
>
>
> Regards,
>
>
>
> Ruediger
>
>
>
> *Von:* Luca Muscariello <[e-mail]>
> *Gesendet:* Dienstag, 28. September 2021 16:50
> *An:* Geib, Rüdiger <[e-mail]>
> *Cc:* Jonathan Morton <[e-mail]>; tsvwg IETF list <
> [e-mail]>
> *Betreff:* Re: [tsvwg] start of WGLC on L4S drafts
>
>
>
> Hi Ruediger,
>
>
>
> off topic, just FYI.
>
>
>
>
>
> On Tue, Sep 28, 2021 at 4:19 PM <[e-mail]> wrote:
>
> Jonathan,
>
> We are working with vendors shipping their own HW (or partially do so),
> and they also integrate HW on their own. QoS isn't the key driver, when
> these systems are specified. VoQ, "per flow treatment" and fair scheduling
> are part of the discussions we have, but on backbone scale. PIE has been
> created by Cisco. That doesn't mean, it has or will be deployed in all
> their gear. Linecard throughput is in Tbit/s range on backplane,
> dequeue-speed is on Gbit/s range. Buffers have to cope with that, when they
> operate fair per flow per queue features.
>
>
>
> Even if you're talking about a different routing set of products, the ASIC
> that is used in the Cisco Nexus 9000 series
>
>
>
>
> https://www.ciscolive.com/c/dam/r/ciscolive/us/docs/2019/pdf/CTHDCN-2301.pdf
>
>
>
> is using flow aware technologies, like AFD and DPP that scale very well in
> hardware
>
> as they need no egress scheduling.
>
>
>
> Best
>
> Luca
>