We use five for our subgroups and two second traffic runs to obtain single
variate gaussians for control charts.

A gaussian measurement is ok for capacity but not for latency or
responsiveness. There, a kolmogorov-smirnov against an ideal or a
comparison set can be used. Or the 99% per the tail of a non-linear
distribution - though that may not be sufficiently descriptive or
correlated for a QoE metric

Then if the transport has a 3WHS, then that is useful. So are flow
completion timed. And typically the duration of slow start is useful.

Bob

On Mon, Dec 9, 2024, 10:53 AM Livingood, Jason via Nnagain <
nnagain@lists.bufferbloat.net> wrote:

> >w/a 100ms baseline. Hilarious. Sad.
>
> [SM] This is not all that terrible for minimal RTT measurements, it does
> however do zilch for latency under load. The 100ms is also not that
> terrible, given that the US is large, think test point in Anchorage
> reflector in Miami...
>
> [JL] The 100ms was initially put there to exclude satellite internet from
> being eligible (prior to LEO operators coming onto the scene).
>
> [JL] Some concerns I noticed in the document:
>
> 1. They want 10% of homes tested in Sec. 3.2, which seems to be an
> extremely large percentage - well above typical statistical significance -
> the FCC MBA only needed >30 to be valid nationally. Amusingly, the example
> they cite in Sec. 3.3 works out to 5% - so they are not internally
> consistent.
> 2. In Sec. 3.4, expecting the ISP to temporarily upgrade subs to the
> highest tier to run tests and then downgrade them again does not make
> practical sense – for example they suggest that rather than randomly
> selecting from users in the highest tier that instead you must select from
> all tiers and then upgrade those not on the highest tier.
> 3. In Sec. 3.9, it seems a bit too proscriptive on the (IXP) server
> locations – could be simplified to a regional peering location of the ISP
> network to allow for flexibility.
> 4. In Sec. 3.10, similarly proscriptive for example by defining the
> duration of a test as 10-15 seconds – what if they test can be completed
> just as accurately in 9 seconds? ISPs should have latitude to configure
> these tests & the state of the art is moving fast.
>
>
>
>
> _______________________________________________
> Nnagain mailing list
> Nnagain@lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/nnagain
>