Re: [rmcat] Still looking for reviews of draft-ietf-rmcat-nada

[Ingemar Johansson S <[e-mail]>]

Hi Xiaoqing + others

To your ECN question... This is tricky, if I understand things right you tie NADA to a specific ECN marking algorithm. For instance CoDel will display a completely different ECN marking behavior. The snag is that with classic ECN marking like this you can only indicate the existence of congestion. It is AFAIK possible to accurately indicate the degree of congestion. 
L4S is a different story as it allows to indicate the degree of congestion.

/Ingemar 


> -----Original Message-----
> From: Xiaoqing Zhu (xiaoqzhu) [mailto:[e-mail]]
> Sent: den 20 juli 2017 19:48
> To: Flohr, Julius <[e-mail]>; Sergio Mena de la Cruz (semena)
> <[e-mail]>
> Cc: Anna Brunstrom <[e-mail]>; Ingemar Johansson S
> <[e-mail]>; [e-mail]; draft-ietf-rmcat-
> [e-mail]
> Subject: Re: [rmcat] Still looking for reviews of draft-ietf-rmcat-nada
> 
> Catching up on this thread.
> 
> Thanks much to Julius, Ingmar, and Safiqul for reviewing this draft and
> providing your input.
> 
> Regarding Ingmar's comments:
> 
> * Page 8, computation of d_tilde and the risk of being locked in the loss-
> based mode due to non-linear warping:  fully agree that this is a limitation of
> the algorithm we should acknowledge in the draft.  In the meanwhile would
> like to work with Julius to look into specific example failure cases, and
> hopefully derive further insights on how to better tune the algorithm
> parameters.  Will update both our ns3-rmcat implementation and NADA draft
> after that investigation.
> 
> * Section 5.1.2 : Estimation of p_mark : as I understand your comments point
> to the fact that different ECN marking behaviors may occur and some (e.g.,
> L4S) may not interact well with NADA.  It is true that we have not tested
> NADA extensively against different variants of ECN markings --- the one's
> we've tried before are RED (marking instead of dropping).  One way to help
> clarify that will be to point to our Appendix A.2 as the specific ECN marking
> behavior we expect, in Sec. 5.1.2.  Do you think that will address your
> concern?
> 
> And, thanks to Julius for sharing with us issues you've observed in your
> implementation efforts.  We've set up a separate email thread to meet up
> remotely and go over that in greater technical details.
> 
> Question back to the Chairs:  our plan now is to update both our draft (to
> address both sets of Ingmar's comments) and corresponding ns3 open-
> source implementation, within a month.  Procedure-wise, are we expecting
> the updated version (-05) to go through WGLC again at that point?  Just trying
> to understand what to expect next.
> 
> Thanks,
> Xiaoqing
> 
> 
> 
> ________________________________________
> From: Flohr, Julius <[e-mail]>
> Sent: Wednesday, July 19, 2017 6:19 AM
> To: Sergio Mena de la Cruz (semena)
> Cc: Anna Brunstrom; Ingemar Johansson S; [e-mail]; draft-ietf-rmcat-
> [e-mail]
> Subject: Re: [rmcat] Still looking for reviews of draft-ietf-rmcat-nada
> 
> Hi,
> 
> > On 19. Jul 2017, at 09:01, Sergio Mena <[e-mail]> wrote:
> >
> >
> > As for Julius's tests, we thank Julius for the effort put into it, and would like
> to see how his results compare to our latest ones (see Slides presented by
> Xiaoqing in April's RMCAT interim).
> >
> 
> I haven't done too much testing, but as far as I can tell there is an issue when
> the bottleneck bandwidth is smaller than RMAX:
> If you have a competition scenario between NADA and a loss based flow, it
> seems that nada gets stuck in the competitive mode because itself creates
> queueing delay that is larger than QTH.
> 
> I actually made a little mistake when I said the problem has gotten worse
> compared to the last version of the draft. The issue actually has been
> introduced in version 03 with the introduction of TEXPLOSS. Before, the draft
> was very unclear and stated that one should warp the congestion signal IFF
> packet loss is present, but without stating a time span.
> 
> Therefore the implementation just used the knowledge it had from the
> recent observation window LOGWIN (500 ms). That worked really well
> because it helped the algorithm to return to the normal mode of operation
> more easily, because after that timespan it would stop warping regardless
> whether there is lots of queueing delay present or not.
> 
> I don't know if this is an issue with my implementation or if there is an actual
> issue with the algorithm, but this is what I have observed.
> 
> I'll contact Xiaoqing and Sergio directly so we can work together on this issue
> and post our findings to this mailing list directly.
> 
> Regards,
> 
> Julius
> 
> 
> =====================================
> Julius Flohr, M.Sc.
> University of Duisburg-Essen
> Computer Networking Technology Group
> 
> Room SL-408
> Schützenbahn 70
> D-45326 Essen/Germany
> 
> E-Mail:  [e-mail]
> Phone: [phone]
> Fax:     [phone]
> =====================================