Re: [tcpm] WGLC for draft-ietf-tcpm-alternativebackoff-ecn-06

Thanks Gorry, Michael,

so to sum it up - a CC should document a safe range of tunables, and the ABE 
variant would pick values within these intervals that may be slightly more 
agressive, vs. loss based.

The guidance on the safe interval of Cubic is noted, and I'll keep that in 
mind!

Best regards,
   Richard

----- Original Message ----- 
From: "G Fairhurst" <gorry@erg.abdn.ac.uk>
To: "Michael Welzl" <michawe@ifi.uio.no>
Cc: "Richard Scheffenegger" <rs.ietf@gmx.at>; <tcpm@ietf.org>
Sent: Monday, March 19, 2018 5:46 PM
Subject: Re: [tcpm] WGLC for draft-ietf-tcpm-alternativebackoff-ecn-06

See end.

On 19/03/2018, 16:34, Michael Welzl wrote:
> Hi,
>
>
>> On Mar 19, 2018, at 3:31 PM, Richard Scheffenegger <rs.ietf@gmx.at 
>> <mailto:rs.ietf@gmx.at>> wrote:
>>
>>
>> Hi Michael,
>>
>> Just thinking that the difference between loss vs. ECN cwn reduction 
>> might be defined as a relative value of each other, rather than an 
>> absolute;
>
> Ah, got it.
>
>
>> In my example, Cubic on FreeBSD used beta 0.8, while Cubic Linux uses 
>> 0.7; with the research showing a sensible Cubic ECN reaction on Linux 
>> being 0.85, correct? (~50% closer to 1).
>
> Correct - though beta_ecn = 0.8 also worked fine for Cubic; 0.85 just 
> seemed to be a slightly better choice in our experiments.
>
>
>> So when using Cubic w/ beta 0.8 (as FreeBSD), would a sensible value for 
>> Cubic ECN there a value of 0.9?
>
> - no, I don’t think it works like this. Whether a value for backoff with 
> ECN is good or bad depends on the overall algorithm behavior, including 
> its increase behavior - not just on its loss backoff factor.
>
+1

The range we put in the ID is [0.7,0.85]. If you are hunting for a
suggestion, It would be that 0.85 is reasonable choice for a safe
reduction for a cubic experiment.

Gorry
> Cheers,
> Michael
>
>
>
> _______________________________________________
> tcpm mailing list
> tcpm@ietf.org
> https://www.ietf.org/mailman/listinfo/tcpm

---
This email has been checked for viruses by AVG.
http://www.avg.com