Re: [tcpm] "David's proposal" for cc class terms (was RE: [tsvwg] L4S status tracking)

[Bob Briscoe <[e-mail]>]

David, Wes,

On 13/11/2019 17:43, Wesley Eddy wrote:
> On 11/13/2019 11:15 AM, Black, David wrote:
>>
>> In more detail than I explained earlier: in this forum
>>
>> We are dealing with two classes of congestion controls.  For lack of 
>> better terms the following class names are based on what the 
>> transport protocol throughput is proportional to where ‘p’ is the 
>> loss and/or congestion marking probability:
>>
>> - 1/sqrt(p)-class congestion controls: Includes most existing TCP 
>> congestion control algorithms, e.g., NewReno, CUBIC.
>>
1/sqrt(p) only describes the worst case that the coupled dualq AQM is 
designed to allow for. It includes cubic in it's reno-friendly regime, 
as well as beyond that. It includes compound and a whole host of 
slightly more aggressive congestion control variants than Reno, as well 
as ABE variants of each. It includes the proprietary congestion 
controllers used in Skype, Hangouts, Facetime, Zoom, Whereby, etc. It 
includes rmcat congestion controllers.

So I wouldn't want to use a name that implied everything in this class 
obeyed (or had to obey) a 1/sqrt law.


Bob


>> - 1/p-class congestion controls: Includes DCTCP congestion control.
>>
>> Keep in mind that p is a probability that is usually << 1 when 
>> expressed as a decimal, e.g., p=0.01 represents a 1% loss/marking rate.
>>
> "Scalable" is the term we've been using in the L4S documents for 1/p, 
> defined in l4s-arch as:
>
>    Scalable Congestion Control:  A congestion control where the packet
>       flow rate per round trip (the window) is inversely proportional to
>       the level (probability) of congestion signals.  Then, as flow rate
>       scales, the number of congestion signals per round trip remains
>       invariant, maintaining the same degree of control.  For instance,
>       DCTCP averages 2 congestion signals per round-trip whatever the
>       flow rate.
>
>
>

-- 
________________________________________________________________
Bob Briscoe                               http://bobbriscoe.net/