Re: [tsvwg] ecn-encap-guidelines reframing section

[Markku Kojo <[e-mail]>]

Hi all,

just one additional comment on using Reno formula an 'p' in it.

On Wed, 24 Mar 2021, Jonathan Morton wrote:

> A quick aside:
>
>> On 24 Mar, 2021, at 11:48 am, Bob Briscoe <[e-mail]> wrote:
>>
>> {Note 1} For instance, the average proportion of marked packets is 'p' in the well known Reno formula,
>>     cwnd_avg = sqrt(3/2p)
>
> This assumes that "p" is a uniform probability; that the likelihood that one packet is marked is independent of whether its neighbours are marked.  A mechanism that refers specifically to marking consecutive packets breaks that assumption.
>
> Reno's actual behaviour depends more accurately on the relationship between the interval between RTTs containing at least one mark, and the actual RTT.

Right. Translating this to Reno formula jargon: p is not the average 
proportion of marked packets but the average proportion of congestion 
signals, where packets marked on the same RTT are considered as a single 
congestion signal. Thereby, p defines the average interval between two 
marks in different RTTs (after removing any additional marks on the same 
RTT). Once we know the path end-to-end bandwidth-delay product and the 
average amount of queuing defined for the bottleneck AQM, we can 
calculate the average RTT to be used in the "full" formula (the target 
queue length and the drop/mark method the AQM applies affects the queuing 
delay and thereby e2e RTT).

And, when thinking of a typical TCP behavior in CA, the typical interval 
between two pkts that an AQM marks for a flow is several RTTs (unless 
there is relatively heavy congestion due to notable number of competing 
flows).

/Markku

> - Jonathan Morton