Re: [tsvwg] new tests of L4S RTT fairness and intra-flow latency: defaults ready for testing

[Jonathan Morton <[e-mail]>]

> On 18 Nov, 2020, at 8:58 am, Ingemar Johansson S <[e-mail]> wrote:
> 
> The difference does not look extreme to me.

It's a persistent and stable 2:1 throughput ratio for about 5-10 seconds after flow startup, followed by a reasonably sharp switch to near parity.  I didn't say it was extreme, only confirmed for Sebastian's benefit that it was clearly noticeable in the data.

> Can’t guarantee it but I suspect that you’ll find similar anomalies with e.g. rate based congestion controls. We had all sorts of interflow fairness issues in the RMCAT work that ended up in RFCs for NADA and SCReAM and L4S was not on the table then. 
>  
> Besides this, I don’t want to kill this discussion, it is very useful. But… it delves into congestion control research and perhaps ICCRG is a better venue for this ?

I'm inclined to agree, except that the RTT independence requirement appears in TSVWG adopted drafts, and TCP Prague is presently the only CC available to test compliance.

What we're showing is that DualPI2 is not sufficient by itself to implement that particular requirement, and arguably makes matters worse than either the typical status quo (dumb FIFO) or the present state of the art (FQ-AQM or AF-AQM).  If all discussion of TCP Prague is to move to ICCRG, then what we will discuss in TSVWG is DualPI2's many shortcomings instead.

 - Jonathan Morton