Re: [tcpm] [EXTERNAL] Re: Linux doesn’t implement RFC3465

On Mon, Aug 2, 2021 at 6:12 PM Yuchung Cheng <ycheng@google.com> wrote:

>
>
> On Mon, Aug 2, 2021 at 5:53 PM Neal Cardwell <ncardwell=
> 40google.com@dmarc.ietf.org> wrote:
>
>>
>>
>> On Mon, Aug 2, 2021 at 8:46 PM Praveen Balasubramanian <pravb=
>> 40microsoft.com@dmarc.ietf.org> wrote:
>>
>>> In experiments a few years ago on DC networks, values over L=8 resulted
>>> in a noticeable increase in packet drops and retransmissions (without
>>> pacing). Windows TCP has been using L=8 for many years now. If we do want
>>> to specify a fallback L value for implementations that cannot pace, my
>>> suggestion would be to use the value 8.
>>>
>>>
>>>
>>> Neal, are there cases where Linux is or can be deployed with infinite L
>>> and no pacing?
>>>
>>
>> Yes, "infinite L and no pacing" is the default behavior for Linux TCP,
>> starting in 2013 for slow-start and then starting in 2015 for congestion
>> avoidance.
>>
> To be more clear: both fq_pacing and TCP pacing have been disabled by
> default in Linux upstream. We do not know how much Linux senders enable
> them today besides the Google servers.
>
> Regarding L = 8, to avoid another round of why or why not. We could say
> inf-L causes line-rate burst up to the stretched ACK degree so put a
> comfortable L if you prefer, then mention implementation practice like
> yours. At the end of the day it's ad-hoc (or "art") and subject to change.
> It might be sensible to cap at cwnd to disincentivize receivers /
> middle-boxes bunching up 10 rounds of ACKs.
>
Sorry please ignore my previous message about the cwnd cap. It is
completely unnecessary -- since with ack-clocking and appropriate counting,
a correct sender would never release more than a cwnd-worth of data. I was
imagining the multiple application-limited burst could let the receiver
keep holding up ACKs, but that can never exceed a cwnd worth of data.

>
>>
>
>> Yuchung pasted the URLs for the exact Linux commits above, which are:
>>
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=9f9843a751d0a2057f9f3d313886e7e5e6ebaac9
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=9cd981dcf174d26805a032aefa791436da709bee
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=c22bdca94782f05b9337d8548bde51b2f38ef17f
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=814d488c61260521b1b3cc97063700a5a6667c8f
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e73ebb0881ea5534ce606c1d71b4ac44db5c6930
>>
>> But I understand that not everyone is in a position to read GPL-licensed
>> code. :-)
>>
>> best regards,
>> neal
>>
>>
>>
>>>
>>>
>>> *From:* tcpm <tcpm-bounces@ietf.org> *On Behalf Of * Neal Cardwell
>>> *Sent:* Monday, August 2, 2021 4:18 PM
>>> *To:* Vidhi Goel <vidhi_goel@apple.com>
>>> *Cc:* Extensions <tcpm@ietf.org>; Mark Allman <mallman@icir.org>
>>> *Subject:* [EXTERNAL] Re: [tcpm] Linux doesn’t implement RFC3465
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Mon, Aug 2, 2021 at 7:02 PM Vidhi Goel <vidhi_goel@apple.com> wrote:
>>>
>>>
>>>
>>> On Mon, Aug 2, 2021 at 3:37 PM Mark Allman <mallman@icir.org> wrote:
>>>
>>>
>>> > The fact is that Linux CC has long moved to infinite L since 2031,
>>>
>>> So, if our experience is with L=\infinity and it is demonstrably OK
>>> why don't we say *THAT* instead of "make L=5 or L=10"?  I would
>>> submit that it makes more sense to leverage experience than it does
>>>
>>> to make things up.
>>>
>>> +1
>>>
>>>
>>>
>>> Yes, I agree that would be a great approach to take.
>>>
>>>
>>>
>>> So, we are saying it is fine to ignore L completely and simply increase
>>> cwnd by bytes_acked during slow start? And if this causes large bursts to
>>> be sent out (when an implementation doesn’t do pacing), that is fine?
>>>
>>>
>>>
>>> Yes, I think that is the proposal on the table, and it sounds good to me.
>>>
>>>
>>>
>>> A rationale would be:
>>>
>>>
>>>
>>> (1) Implementations SHOULD pace (RFC 7661).
>>>
>>>
>>>
>>> (2) Implementations that don't pace will generally be causing large
>>> bursts for many different reasons anyway (data and/or ACK aggregation in
>>> the network or end hosts), restart from idle,...) so having a constant L
>>> does not provide enough protection from bursts to justify the cost in
>>> reduced performance (in the form of slower slow-start). In support of this,
>>> experience with this as the default behavior in Linux TCP over the
>>> 2013-2021 period suggests this works well enough in practice.
>>>
>>>
>>>
>>> neal
>>>
>>>
>>>
>>>
>>>
>> _______________________________________________
>> tcpm mailing list
>> tcpm@ietf.org
>> https://www.ietf.org/mailman/listinfo/tcpm
>>
>