Re: [tcpm] Changes to draft-ietf-tcpm-hystartplusplus after WGLC

[Praveen Balasubramanian <[e-mail]>]

We will add text that disambiguates what the default was for.

On Wed, Dec 7, 2022 at 10:01 AM Gorry Fairhurst <[e-mail]>
wrote:

> On 06/12/2022 15:00, Neal Cardwell wrote:
>
>
>
> On Tue, Dec 6, 2022 at 4:09 AM Gorry Fairhurst <[e-mail]>
> wrote:
>
>> On 05/12/2022 20:07, Neal Cardwell wrote:
>> > On Mon, Dec 5, 2022 at 4:48 AM Gorry Fairhurst <[e-mail]>
>> wrote:
>> >> On 05/12/2022 09:03, [e-mail] wrote:
>> >>>> On 29. Nov 2022, at 19:31, Neal Cardwell <ncardwell=
>> [e-mail]> wrote:
>> >>>>
>> >>>> On Tue, Nov 29, 2022 at 7:05 AM <[e-mail]> wrote:
>> >>>> Dear all,
>> >>>>
>> >>>> during the AD review, some changes have been made to
>> draft-ietf-tcpm-hystartplusplus-09
>> >>>> in relation with RFC 3465 (TCP Congestion Control with Appropriate
>> Byte Counting (ABC))
>> >>>> to avoid referencing an informational document from a standards
>> track document.
>> >>>>
>> >>>> These changes can be reviewed using:
>> >>>>
>> https://www.ietf.org/rfcdiff?url1=https://www.ietf.org/archive/id/draft-ietf-tcpm-hystartplusplus-09.txt&url2=https://www.ietf.org/archive/id/draft-ietf-tcpm-hystartplusplus-11.txt
>> >>>>
>> >>>> Please indicate within a week, whether you are fine with these
>> changes or provide
>> >>>> some comments.
>> >>>>
>> >>>> Thanks for posting the diff!
>> >>>>
>> >>>> Regarding this part of section 5, "Deployments and Performance
>> Evaluations":
>> >>>>
>> >>>> The original Hystart has been default-enabled
>> >>>> for all TCP connections in the Linux operating system using the
>> >>>> default congestion control module CUBIC ([RFC8312]) for a decade with
>> >>>> L = infinity with pacing enabled.
>> >>>> The "with pacing enabled" part looks inaccurate. Linux TCP thus far
>> has not enabled pacing by default. This passage would be correct if
>> modified to read: "for a decade with L = infinity with pacing disabled by
>> default."
>> >>> Now I'm confused. I thought the reason for "SHOULD use L=Infinity
>> when pacing enabled" was that
>> >>> this setting is used in Linux for years as the default. So what is
>> the reason now?
>> >>>
>> >>> Best regards
>> >>> Michael (as an individual)
>> >> I'm also now wondering what is the case:
>> >>
>> >> To me at least, the pacing text describes a way we can more concretely
>> >> say that that bursts were mitigated (albeit without saying how to
>> pace).
>> >>
>> >> Is the use of the word "default" in some way confusing?
>> > IMHO it is not confusing. :-) If someone could outline how it is
>> > confusing, that could be useful.
>>
>> When I first read this, I thought I read  "when pacing is enabled",
>> which seems consistent with what you say, but the ID did not actually
>> read that way. The text says:
>>
>> "The original Hystart has been default-enabled
>>     for all TCP connections in the Linux operating system using the
>>     default congestion control module CUBIC ([RFC8312]) for a decade with
>>     L = infinity with pacing enabled."
>>
>> - my question was therefore about whether Linux distributions have then
>> enabled pacing by default ?
>>
>
> I mentioned this in my first e-mail in this thread (from Nov 29):
>
>    The "with pacing enabled" part looks inaccurate. Linux TCP thus far
>    has not enabled pacing by default. This passage would be correct if
>    modified to read: "for a decade with L = infinity with pacing disabled
> by default."
>    ...
>    ...the "pacing enabled" part is only true of Linux installations that
> use
>    the "fq" qdisc, which is not the default for the Linux kernel or any
> major
>    Linux distribution that I'm aware of ("fq_codel" is increasingly common
>    among major Linux distributions, but it does not implement pacing).
>
> If you need more detail: I just checked, and of the most recent versions
> of the most popular Linux distributions the default qdiscs used are:
>
>    + Ubuntu 22.04 and CentOS Stream 9 use fq_codel
>    + Debian 11 and RHEL 7 use pfifo_fast
>
> Neither fq_codel nor pfifo_fast implement pacing. So this seems to confirm
> that pacing is still not enabled by default on major Linux distributions.
>
> neal
>
> >> I was hoping
>> >> there had been experience in some specific networks using "L=Infinity
>> >> when pacing was enabled",
>> > Yes, there is extensive experience with large-scale deployment of this
>> > approach (paced CUBIC with Hystart and L = infinity.), including all
>> > YouTube and Google TCP traffic (both internally and over the public
>> > Internet) in the 2013-2016 era.
>> >
>> Good - that is what I thought about these using pacing.
>> >> and was expecting that the spec only permitted
>> >> when bursts were avoided, e.g., pacing was enabled?
>> > Yes, the Hystart++ spec already does this; it says:
>> >
>> >    L = infinity if paced, L = 8 if non-paced
>> >
>> > best regards,
>> > neal
>>
>> Best wishes,
>>
>> Gorry
>>
>>
>> >> Gorry
>> >>
>> >>
>> >>>> To be clear, the algorithm variant mentioned here (Linux TCP with
>> CUBIC Hystart and L=infinity and pacing enabled) has had plenty of air
>> miles, including YouTube and Google TCP traffic in the 2013-2016 era. But
>> the "pacing enabled" part is only true of Linux installations that use the
>> "fq" qdisc, which is not the default for the Linux kernel or any major
>> Linux distribution that I'm aware of ("fq_codel" is increasingly common
>> among major Linux distributions, but it does not implement pacing).
>> >>>>
>> >>>> The rest of the update looks good to me.
>> >>>>
>> >>>> Thanks!
>> >>>>
>> >>>> best regards,
>> >>>> neal
>> >>>>
>> >>>> Best regards
>> >>>> Michael_______________________________________________
>> >>>> tcpm mailing list
>> >>>> [e-mail]
>> >>>> https://www.ietf.org/mailman/listinfo/tcpm
>> >>>> _______________________________________________
>> >>>> tcpm mailing list
>> >>>> [e-mail]
>> >>>> https://www.ietf.org/mailman/listinfo/tcpm
>> >>> _______________________________________________
>> >>> tcpm mailing list
>> >>> [e-mail]
>> >>> https://www.ietf.org/mailman/listinfo/tcpm
>> >>
>>
>> OK, I think I udnerstand and this is close to what I expected, thanks for
> explaining.
>
> My comment arose from the current text that read:
>
> "The original Hystart has been default-enabled
>
>    for all TCP connections in the Linux operating system using the
>    default congestion control module CUBIC ([RFC8312]) for a decade with
>    L = infinity with pacing enabled."
> - Which I took as saying that pacing was by default enabled, I still read it this way.
>
> Do you think we can rephrase this, depending on what was the case?
>
> For example:
>
>    "There has been over a decade of experience using the original Hystart in
>      Linux deployment with pacing that default-enabled the original Hystart
>      with L = infinity for all TCP connections using the
>      default congestion control module, CUBIC [RFC8312]."
>
> Or something similar that clearly says where the default was for the pacing?
>
> Gorry
>
>
>
>
>
>
>
>