BBR parameters

[Habte Kassaye]

Hi,

I wanted to experiment on  BBR parameters like probeBW frequency and duration, but I do not see any modinfo parameter list for the tcp_bbr module.
Is it not possible to change/access those parameters through the kernel as module parameters? 
If it is not possible to change those values could you please tell me why some of the parameters, such as the 6-10RTT probeBW interval and 10 sec probeRTT interval, are set to those values?

Thanks,
Habte

[Beyers Cronje]

I can not comment on how to change the parameters, but the internet draft does cover why some of the parameters are set to their specific values. See for instance https://tools.ietf.org/html/draft-cardwell-iccrg-bbr-congestion-control-00#section-4.3.5

[Habte Kassaye]

 Thanks for the suggestion. I just had a quick look and it explains the probeRTT selection well. I still don't understand the selection of 8 phases for the probeBW  cycle though. I'll see if it is explained in the draft.

[Neal Cardwell]

Yes, the Internet Draft has the most in-depth discussion of the rationale for the various aspects of the BBR design:

   https://tools.ietf.org/html/draft-cardwell-iccrg-bbr-congestion-control

The BBR.BtlBw estimator and ProbeBW and gain cycling parameters are discussed here:

  https://tools.ietf.org/html/draft-cardwell-iccrg-bbr-congestion-control-00#section-4.1.1

  https://tools.ietf.org/html/draft-cardwell-iccrg-bbr-congestion-control-00#section-4.3.4

The BBR.RTprop estimator and ProbeRTT parameters are discussed here:

  https://tools.ietf.org/html/draft-cardwell-iccrg-bbr-congestion-control-00#section-4.1.2

  https://tools.ietf.org/html/draft-cardwell-iccrg-bbr-congestion-control-00#section-4.3.5

As you noticed, in the Linux TCP BBR code there is no module parameter interface for BBR. This is intentional. The parameters for BBR are "design" parameters, rather than "configuration" parameters. These parameters are not intended to be configured by system administrators. They were all chosen carefully, with broad-based long-term testing, as part of designing the algorithm itself. And modifying them can lead to serious performance and fairness problems. The parameters are designed to "just work" across a broad range of environments, right out of the box, without tedious and error-prone manual tuning. That is why the Linux TCP BBR kernel module has no module parameters.

Developers in the open source or research community who are working on improving the algorithm will necessarily be compiling their own test kernels, and in that development process can temporarily add to their source tree whatever module parameters make sense for their (careful) internal testing and experimentation.

cheers,

neal

--
You received this message because you are subscribed to the Google Groups "BBR Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email to bbr-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Habte Kassaye]

So if I understand correctly the "ProbeBW gain cycle" is limited to 8 phases so that a cycle is included in a "BtlBwFilterLen" interval, which is 10 round trips long. The "BtlBwFilterLen" is set to 10 to allow quick response/adaptation while avoid transient fluctuations that were observed while testing for these parameters. Does this mean that these parameters could change in the future if the nature of the the noise/short fluctuations changes?

Thanks,
Habte

To unsubscribe from this group and stop receiving emails from it, send an email to bbr-dev+u...@googlegroups.com.

[Neal Cardwell]

Yes, my sense is that BBR algorithm parameters could change at some point in the future if broad testing supports this.

neal

To unsubscribe from this group and stop receiving emails from it, send an email to bbr-dev+unsubscribe@googlegroups.com.