Re: [tcpm] draft-ietf-tcpm-prr-rfc6937bis-03: set cwnd to ssthresh exiting fast recovery?

[Neal Cardwell <[e-mail]>]

On Wed, Aug 9, 2023 at 11:39 AM Yuchung Cheng <[e-mail]> wrote:

> We can add "Upon exiting recovery, cwnd /SHOULD/ be set to ssthresh" with
> some performance rationale, given that RFC6675 and existing PRR
> implementation already do so.
>

This sounds very good to me. Yoshifumi, how does that sound?

thanks,
neal


> Note that RFC5681 Sec 4.3 has related wording on cwnd exiting recovery:
> "Finally, after all loss in the given window of segments has been
> successfully retransmitted, cwnd MUST be set to no more than ssthresh and
> congestion avoidance MUST be used to further increase cwnd."
>
> Why not MUST: it's not strictly necessary because it won't break TCP or
> make network unstable. It's important for congestion control to determine
> the cwnd after recovery. It has the implication to induce a large burst if
> cwnd >> pipe as mentioned in the end of Section 5 in RFC6675.
> Why not MAY: lacking so has major performance implications in various
> cases as discussed in this thread
>
> How does that sound?
>
> On Wed, Aug 9, 2023 at 8:10 AM Neal Cardwell <[e-mail]> wrote:
>
>>
>>
>> On Wed, Aug 9, 2023 at 11:05 AM Neal Cardwell <[e-mail]>
>> wrote:
>>
>>> Hi Yoshifumi,
>>>
>>> You are correct that draft-ietf-tcpm-prr-rfc6937bis-04 does not
>>>  incorporate the suggestion in this thread to have a "cwnd = ssthresh" step
>>> at the end of fast recovery. My sense was that this was because we had not
>>> come to a conclusion / resolution of this question in this thread. :-)
>>>
>>> I would still argue that it's important for PRR to set cwnd = ssthresh
>>> at the end of recovery. Without setting cwnd = ssthresh at the end of
>>> recovery, cwnd could end recovery far below ssthresh, leading to unusably
>>> terrible performance; performance that would be far worse than RFC 6675
>>> recovery (which simply sets cwnd = ssthresh at the start of recovery).
>>>
>>> The Linux TCP PRR has had this cwnd = ssthresh step at the end of
>>> recovery since the original PRR implementation in 2011:
>>>
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a262f0cdf1f2916ea918dc329492abb5323d9a6c
>>>
>>
>> And FWIW it sounds like from Randall Stewart's earlier post on this
>> thread ("when we exit recovery we set cwnd to ssthresh") that FreeBSD TCP
>> PRR also has the same  cwnd = ssthresh step at the end of recovery that
>> Linux TCP PRR has.
>>
>> I would suspect that Microsoft TCP PRR has a similar step; I've CC-ed
>> some folks who may be able to shed light on that.
>>
>> neal
>>
>>
>>
>>> best regards,
>>> neal
>>>
>>>
>>>
>>>
>>> On Wed, Aug 9, 2023 at 3:16 AM Yoshifumi Nishida <[e-mail]>
>>> wrote:
>>>
>>>> Hi Yuchung,
>>>>
>>>> Thanks for the response.
>>>> I just would like to check one thing.
>>>> In my understanding, Neal's suggestion here was to adjust cwnd to
>>>> ssthresh at the end of recovery.
>>>> But, I cannot find the statement or logic for such adjustment. Does
>>>> this mean we decided there's no adjustment at the end of recovery? Or, am I
>>>> missing something?
>>>>
>>>> Thanks,
>>>> --
>>>> Yoshi
>>>>
>>>> On Tue, Aug 8, 2023 at 2:34 PM Yuchung Cheng <[e-mail]> wrote:
>>>>
>>>>> Hi Yoshifumi,
>>>>>
>>>>> That part is how the "RecoverFS" state variable is calculated in the
>>>>> draft. See the diff of 03/04 on Section 5 and 6 regarding "RecoverFS" state
>>>>> variable definition and computation.
>>>>> https://author-tools.ietf.org/iddiff?url2=draft-ietf-tcpm-prr-rfc6937bis-04
>>>>>
>>>>> Does that make sense?
>>>>>
>>>>> On Tue, Aug 8, 2023 at 12:01 AM Yoshifumi Nishida <[e-mail]>
>>>>> wrote:
>>>>>
>>>>>> Hi Yuchung,
>>>>>>
>>>>>> I think you have already updated the draft on the following point
>>>>>> from the discussions in the last WG meeting.
>>>>>> Could you point out which part has been updated? I'm just checking..
>>>>>> Thanks,
>>>>>> --
>>>>>> Yoshi
>>>>>>
>>>>>> On Fri, May 5, 2023 at 11:51 AM Yoshifumi Nishida <[e-mail]>
>>>>>> wrote:
>>>>>>
>>>>>>> Hi Neal,
>>>>>>>
>>>>>>> Yes, I think I understand your point.
>>>>>>> I prefer the current logic in some ways as it's more conservative as
>>>>>>> I think we cannot always presume that queue has been drained at the end of
>>>>>>> recovery.
>>>>>>> But, I also think it may look too conservative.
>>>>>>> I am expecting that the authors provide some insights on this point.
>>>>>>> --
>>>>>>> Yoshi
>>>>>>>
>>>>>>>
>>>>>>> On Tue, May 2, 2023 at 11:31 AM Neal Cardwell <[e-mail]>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Hi Yoshi,
>>>>>>>>
>>>>>>>> You are right that because PRR always sets cwnd to ssthresh at the
>>>>>>>> end of recovery, there will be some cases where with PRR cwnd jumps up
>>>>>>>> drastically at the end of the recovery.
>>>>>>>>
>>>>>>>> However, AFAIK cwnd jumping up drastically, per se, is not a
>>>>>>>> problem. Big bursts of packets going into the network is a problem. And
>>>>>>>> given the dynamics of the alternative loss recovery algorithms (RFC6675 and
>>>>>>>> PRR), both can allow bursts of packets; just in different circumstances:
>>>>>>>>
>>>>>>>> (1) RFC6675: Because RFC6675 sets cwnd once at the start of fast
>>>>>>>> recovery, using (4.2) from RFC6675:
>>>>>>>>
>>>>>>>> ssthresh = cwnd = (FlightSize / 2)
>>>>>>>>
>>>>>>>> ...that means RFC6675 allows big bursts at the moment any loss is
>>>>>>>> detected: any time L packets are lost, the sender can burst L more packets.
>>>>>>>>
>>>>>>>> (2) PRR: PRR is specifically designed to avoid big bursts in
>>>>>>>> response to packet losses; no matter the structure or timing of the losses,
>>>>>>>> PRR only allows a big burst at the end of Fast Recovery after all holes
>>>>>>>> have been plugged, and the algorithm sets cwnd to ssthresh.
>>>>>>>>
>>>>>>>> So in your example ("For example, many packets were lost before
>>>>>>>> entering recovery"), AFAICT RFC6675 can allow a big burst at the beginning
>>>>>>>> of recovery, when the lost packets are detected. AFAICT in this case PRR
>>>>>>>> can allow a burst of packets at the end of recovery when it sets cwnd to
>>>>>>>> ssthresh, but at least at this point the bottleneck queue has potentially
>>>>>>>> drained somewhat.
>>>>>>>>
>>>>>>>> Please let me know if that analysis misses something important. :-)
>>>>>>>>
>>>>>>>> Thanks!
>>>>>>>> neal
>>>>>>>>
>>>>>>>>
>>>>>>>> On Mon, May 1, 2023 at 5:22 PM Yoshifumi Nishida <
>>>>>>>> [e-mail]> wrote:
>>>>>>>>
>>>>>>>>> Hi Randall,
>>>>>>>>>
>>>>>>>>> I might miss something, but here's what I've thought..
>>>>>>>>> If we lost many packets in a RTT such as the Figure 5 in the
>>>>>>>>> 6937bis draft, I think the window growth during the recovery period will be
>>>>>>>>> bound by PRR-CRB or PRR-SSRB.
>>>>>>>>> Hence, I think the cwnd at the end of recovery can be smaller than
>>>>>>>>> we expect as shown in figure 5.
>>>>>>>>> --
>>>>>>>>> Yoshi
>>>>>>>>>
>>>>>>>>> On Mon, May 1, 2023 at 4:17 AM Randall Stewart <[e-mail]>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Hi Neal and Yoshi:
>>>>>>>>>>
>>>>>>>>>> Neal: So the FreeBSD implementation in rack, like linux, does the
>>>>>>>>>> same exact thing set cwnd to ssthresh at
>>>>>>>>>> exit from recovery.
>>>>>>>>>>
>>>>>>>>>> Yoshi: I don’t see how this would cause cwnd to be larger, since
>>>>>>>>>> at the entry to recovery you set ssthresh = cwnd *  Beta. But
>>>>>>>>>>           maybe I am missing something, can you give an example
>>>>>>>>>> like Neal did below?
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Thanks
>>>>>>>>>>
>>>>>>>>>> R
>>>>>>>>>>
>>>>>>>>>> On May 1, 2023, at 5:32 AM, Yoshifumi Nishida <[e-mail]>
>>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>> Hi Neal,
>>>>>>>>>>
>>>>>>>>>> If we always set cwnd to ssthresh at the end of recovery, I am
>>>>>>>>>> guessing there will be some cases where cwnd jumps up drastically at the
>>>>>>>>>> end of the recovery. For example, many packets were lost before entering
>>>>>>>>>> recovery.  Or, am I missing something?
>>>>>>>>>> --
>>>>>>>>>> Yoshi
>>>>>>>>>>
>>>>>>>>>> On Wed, Apr 19, 2023 at 7:37 PM Neal Cardwell <ncardwell=
>>>>>>>>>> [e-mail]> wrote:
>>>>>>>>>>
>>>>>>>>>>> Working through examples for the
>>>>>>>>>>> "draft-ietf-tcpm-prr-rfc6937bis-03 and RecoverFS initialization" thread
>>>>>>>>>>> this evening, I ran into another potential issue.
>>>>>>>>>>>
>>>>>>>>>>> The Linux TCP implementation of PRR explicitly/directly sets
>>>>>>>>>>> cwnd to ssthresh at the end of fast recovery (in tcp_end_cwnd_reduction()).
>>>>>>>>>>> But this behavior is not in the algorithm in the PRR RFC or draft, at least
>>>>>>>>>>> in the figures in section 6, Algorithms. Maybe it is in the prose somewhere
>>>>>>>>>>> and I missed it; but in that case I'd argue strongly to put this in the
>>>>>>>>>>> figures in section 6, Algorithms.
>>>>>>>>>>>
>>>>>>>>>>> AFAICT in some cases this is strictly necessary to get cwnd to
>>>>>>>>>>> grow to reach ssthresh. Without such a direct step, cwnd could end up far
>>>>>>>>>>> below ssthresh at the end of recovery. Here's an example to illustrate:
>>>>>>>>>>>
>>>>>>>>>>> CC = CUBIC
>>>>>>>>>>>
>>>>>>>>>>> cwnd = 10
>>>>>>>>>>>
>>>>>>>>>>> The reordering degree was estimated to be large, so the
>>>>>>>>>>> connection will wait for more than 3 packets to be SACKed before entering
>>>>>>>>>>> fast recovery.
>>>>>>>>>>>
>>>>>>>>>>> --- Application writes 10*MSS.
>>>>>>>>>>>
>>>>>>>>>>> TCP sends packets P1 .. P10.
>>>>>>>>>>> pipe = 10 packets in flight (P1 .. P10)
>>>>>>>>>>>
>>>>>>>>>>> --- P2..P9 SACKed  -> do nothing
>>>>>>>>>>>
>>>>>>>>>>> (Because the reordering degree was previously estimated to be
>>>>>>>>>>> large.)
>>>>>>>>>>>
>>>>>>>>>>> --- P10 SACKed -> mark P1 as lost and enter fast recovery
>>>>>>>>>>>
>>>>>>>>>>> PRR:
>>>>>>>>>>> ssthresh = CongCtrlAlg() = 7 packets // CUBIC
>>>>>>>>>>> prr_delivered = 0
>>>>>>>>>>> prr_out = 0
>>>>>>>>>>> RecoverFS = snd.nxt - snd.una = 10 packets (P1..P10)
>>>>>>>>>>>
>>>>>>>>>>> DeliveredData = 1  (P10 was SACKed)
>>>>>>>>>>>
>>>>>>>>>>> prr_delivered += DeliveredData   ==> prr_delivered = 1
>>>>>>>>>>>
>>>>>>>>>>> pipe =  0  (all packets are SACKed or lost; P1 is lost, rest are
>>>>>>>>>>> SACKed)
>>>>>>>>>>>
>>>>>>>>>>> safeACK = false (snd.una did not advance)
>>>>>>>>>>>
>>>>>>>>>>> if (pipe > ssthresh) => if (0 > 7) => false
>>>>>>>>>>> else
>>>>>>>>>>>   // PRR-CRB by default
>>>>>>>>>>>   sndcnt = MAX(prr_delivered - prr_out, DeliveredData)
>>>>>>>>>>>          = MAX(1 - 0, 1)
>>>>>>>>>>>          = 1
>>>>>>>>>>>
>>>>>>>>>>>   sndcnt = MIN(ssthresh - pipe, sndcnt)
>>>>>>>>>>>          = MIN(7 - 0, 1)
>>>>>>>>>>>          = 1
>>>>>>>>>>>
>>>>>>>>>>> cwnd = pipe + sndcnt
>>>>>>>>>>>      = 0    + 1
>>>>>>>>>>>      = 1
>>>>>>>>>>>
>>>>>>>>>>> retransmit P1
>>>>>>>>>>>
>>>>>>>>>>> prr_out += 1   ==> prr_out = 1
>>>>>>>>>>>
>>>>>>>>>>> --- P1 retransmit plugs hole; receive cumulative ACK for P1..P10
>>>>>>>>>>>
>>>>>>>>>>> DeliveredData = 1  (P1 was newly ACKed)
>>>>>>>>>>>
>>>>>>>>>>> prr_delivered += DeliveredData   ==> prr_delivered = 2
>>>>>>>>>>>
>>>>>>>>>>> pipe =  0  (all packets are cumuatively ACKed)
>>>>>>>>>>>
>>>>>>>>>>> safeACK = (snd.una advances and no further loss indicated)
>>>>>>>>>>> safeACK = true
>>>>>>>>>>>
>>>>>>>>>>> if (pipe > ssthresh) => if (0 > 7) => false
>>>>>>>>>>> else
>>>>>>>>>>>   // PRR-CRB by default
>>>>>>>>>>>   sndcnt = MAX(prr_delivered - prr_out, DeliveredData)
>>>>>>>>>>>          = MAX(2 - 1, 1)
>>>>>>>>>>>          = 1
>>>>>>>>>>>   if (safeACK) => true
>>>>>>>>>>>     // PRR-SSRB when recovery is in good progress
>>>>>>>>>>>     sndcnt += 1   ==> sndcnt = 2
>>>>>>>>>>>
>>>>>>>>>>>   sndcnt = MIN(ssthresh - pipe, sndcnt)
>>>>>>>>>>>          = MIN(7 - 0, 2)
>>>>>>>>>>>          = 2
>>>>>>>>>>>
>>>>>>>>>>> cwnd = pipe + sndcnt
>>>>>>>>>>>      = 0    + 2
>>>>>>>>>>>      = 2
>>>>>>>>>>>
>>>>>>>>>>> So we exit fast recovery with cwnd=2 even though ssthresh is 7.
>>>>>>>>>>>
>>>>>>>>>>> As noted above, the Linux TCP implementation does not suffer
>>>>>>>>>>> this problem because it explicitly/directly sets cwnd to ssthresh at the
>>>>>>>>>>> end of fast recovery.
>>>>>>>>>>>
>>>>>>>>>>> I would recommend including this cwnd=ssthresh step at the end
>>>>>>>>>>> of recovery in the draft, to ensure that cwnd reaches ssthresh at the end
>>>>>>>>>>> of fast recovery, even in cases like this where there will be insufficient
>>>>>>>>>>> delivered data in fast recovery to allow pipe to incrementally grow to
>>>>>>>>>>> reach ssthresh using PRR-SSRB.
>>>>>>>>>>>
>>>>>>>>>>> Best regards,
>>>>>>>>>>> neal
>>>>>>>>>>>
>>>>>>>>>>> _______________________________________________
>>>>>>>>>>> tcpm mailing list
>>>>>>>>>>> [e-mail]
>>>>>>>>>>> https://www.ietf.org/mailman/listinfo/tcpm
>>>>>>>>>>> <https://www.google.com/url?q=https://www.ietf.org/mailman/listinfo/tcpm&source=gmail-imap&ust=1683538345000000&usg=AOvVaw2cOITQpYcuP_M95396rEmw>
>>>>>>>>>>>
>>>>>>>>>> _______________________________________________
>>>>>>>>>> tcpm mailing list
>>>>>>>>>> [e-mail]
>>>>>>>>>>
>>>>>>>>>> https://www.google.com/url?q=https://www.ietf.org/mailman/listinfo/tcpm&source=gmail-imap&ust=1683538345000000&usg=AOvVaw2cOITQpYcuP_M95396rEmw
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> ------
>>>>>>>>>> Randall Stewart
>>>>>>>>>> [e-mail]
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>