Re: [tsvwg] [Ecn-sane] per-flow scheduling

[Gorry Fairhurst <[e-mail]>]

On 27/06/2019, 08:53, Bless, Roland (TM) wrote:
> Hi,
>
> Am 26.06.19 um 18:31 schrieb David P. Reed:
>
>> The idea that "link utilization" of 100% must be achieved is why we got
>> bufferbloat designed into routers. It's a worm's eye perspective. To
> You are right, but IMHO it is even worse, because it is an
> artefact of the particular loss-based TCP congestion control
> as it was designed and the correspondingly derived "BDP buffer size"
> rule. The loss-based AIMD congestion control is able to keep up the
> utilization then, because it uses the systematically built standing
> queue during its back-off after a packet loss. That essentially keeps
> the sending rate at the same level. However, now we know that avoiding
> queuing delay is important, we can also design better congestion
> controls that do not fill the available buffer capacity up to exhaustion
> and that do not require standing queues to keep up link utilization.
>
> However, as you also wrote, the positive and negative effects of
> existing buffers depend a lot on the particular traffic pattern and this
> has also changed much during the last decades. So I think that revising
> the "buffer size" discussion could be useful...
>
> Aside from that I find the SCE proposal very useful, because it allows
> to provide an additional level of congestion signalling that could
> be used by various congestion control schemes.
>
> Regards,
>   Roland
I know from the work on using ECT(0) with RFC8511, that you can already 
get a lot of benefit from treating ECN as a first indication of 
congestion and loss as a much more serious signal. I may have missed 
something recently. I recall attempts to do 3-level marking of ECNin the 
past,but  when I last looked through the litterature I did not find any 
lasting deployment resulted or modern research  (please send pointers if 
I overlooked something).

I am therefore  rather skeptical that research on SCE is ready  I'd be 
curious to see what extra value is gained from the SCE approach.

Gorry