From: Yuchung Cheng <ycheng@xxxxxxxxxx>
Sent: Thursday, December 17, 2020 2:55 PM
To: Martin Duke <martin.h.duke@xxxxxxxxx>
Cc: Praveen Balasubramanian <pravb@xxxxxxxxxxxxx>; kojo@xxxxxxxxxxxxxx; tcpm@xxxxxxxx;
draft-ietf-tcpm-rack@xxxxxxxx; tuexen@xxxxxxxxxxxxxx; draft-ietf-tcpm-rack.all@xxxxxxxx;
last-call@xxxxxxxx; tcpm-chairs@xxxxxxxx
Subject: Re: [EXTERNAL] Re: [tcpm] Last Call:
<draft-ietf-tcpm-rack-13.txt>(TheRACK-TLPlossdetectionalgorithmforTCP) to Proposed Standard
On Thu, Dec 17, 2020 at 1:59 PM Martin Duke <martin.h.duke@xxxxxxxxx> wrote:
This is good, thanks.
I might also add the sentence "In the absence of PRR, when TCP RACK detects a lost
retransmission it MUST trigger a second congestion response" or something to that
effect. What do you think?
Sure I'd revise a bit to be even more detailed and consistent with the principle in RFC5681
"In the absence of PRR, when RACK-TLP detects a lost retransmission the congestion control
MUST trigger an additional congestion response per the aforementioned principle in RFC5681.
If multiple retransmissions were lost in a window, the congestion control specified in
RFC5681 only reacts once per window, but some other congestion control may want to respond
multiple times differently. The congestion control implementer is advised to
carefully consider this subtle situation."
On Thu, Dec 17, 2020 at 11:36 AM Yuchung Cheng <ycheng@xxxxxxxxxx> wrote:
How about
"9.3. Interaction with congestion control
RACK-TLP intentionally decouples loss detection ...
As mentioned in Figure 1 caption, RFC5681 mandates a principle that
Loss in two successive windows of data, or the loss of a
retransmission, should be taken as two indications of congestion, and
therefore reacted separately. However implementation of RFC6675 pipe
algorithm may not directly account for this newly detected congestion
events properly. PRR [RFCxxxx] is RECOMMENDED for the specific
congestion control actions taken upon the losses detected by RACK-TLP"
To Makku's request for "what's the justification to enter fast recovery".
Consider this example w/o RACK-TLP
T0: Send 100 segments but application-limited. All are lost.
T-2RTT: app writes so another 3 segments are sent. Made to the destination and
triggered 3 DUPACKs
T-3RTT: 3 DUPACK arrives. start fast recovery and subsequent cc reactions to
burst ~50 packets with Reno
In this case any ACK occured before RTO is (generally) considered clock-acked,
and how I understand Van's initial design. This behavior existed decades before
RACK-TLP. RACK-TLP essentially changes the "3-segments by app" to "1-segment by
tcp".
On Thu, Dec 17, 2020 at 10:52 AM Praveen Balasubramanian <pravb@xxxxxxxxxxxxx>
wrote:
I agree that we should have a note in this RFC about congestion
control action upon detecting lost retransmission(s).
From: tcpm <tcpm-bounces@xxxxxxxx> On Behalf Of Martin Duke
Sent: Thursday, December 17, 2020 7:30 AM
To: Markku Kojo <kojo@xxxxxxxxxxxxxx>
Cc: tcpm@xxxxxxxx Extensions <tcpm@xxxxxxxx>;
draft-ietf-tcpm-rack@xxxxxxxx; Michael Tuexen
<tuexen@xxxxxxxxxxxxxx>; draft-ietf-tcpm-rack.all@xxxxxxxx; Last Call
<last-call@xxxxxxxx>; tcpm-chairs <tcpm-chairs@xxxxxxxx>
Subject: [EXTERNAL] Re: [tcpm] Last Call:
<draft-ietf-tcpm-rack-13.txt>(TheRACK-TLPlossdetectionalgorithmforTCP)
to Proposed Standard
Hi Markku,
Thanks, now I understand your objections.
Martin
On Thu, Dec 17, 2020 at 12:46 AM Markku Kojo <kojo@xxxxxxxxxxxxxx> wrote:
Hi,
On Wed, 16 Dec 2020, Martin Duke wrote:
> I spent a little longer looking at the specs more carefully,
and I explained (1)
> incorrectly in my last two messages. P21..29 are not Limited
Transmit packets.
Correct. Just normal the rule that allows sending new data
during fast
recovery.
> However, unless I'm missing something else, 6675 is clear
that the recovery period
> does not end until the cumulative ack advances, meaning that
detecting the lost
> retransmission of P1 does not trigger another MD directly.
As I have said earlier, RFC 6675 does not repeat all congestion
control
principles from RFC 5681. It definitely honors the CC principle
that
requires to treat a loss of a retransmission as a new
congestion
indication and another MD. I believe I am obligated to know
this as a
co-author of RFC 6675. ;)
RFC 6675 explicitly indicates that it follows RFC 5681 by
stating in the
abstract:
" ... conforms to the spirit of the current congestion control
specification (RFC 5681 ..."
And in the intro:
"The algorithm specified in this document is a
straightforward
SACK-based loss recovery strategy that follows the
guidelines
set in [RFC5681] ..."
I don't think there is anything unclear in this.
RFC 6675 and all other standard congestion controls (RFC 5581
and RFC
6582) handle a loss of a retransmission by "enforcing" RTO to
detect it.
And RTO guarantees MD. RACK-TLP changes the loss detection in
this case
and therefore the standard congestion control algorithms do not
have
actions to handle it corrrectly. That is the point.
BR,
/Markku
> Thanks for this exercise! It's refreshed my memory of these
details after working
> on slightly different QUIC algorithms a long time.
>
> On Wed, Dec 16, 2020, 18:55 Martin Duke
<martin.h.duke@xxxxxxxxx> wrote:
> (1) Flightsize: in RFC 6675. Section 5, Step 4.2:
>
> (4.2) ssthresh = cwnd = (FlightSize / 2)
>
> The congestion window (cwnd) and slow start
threshold
> (ssthresh) are reduced to half of FlightSize per
[RFC5681].
> Additionally, note that [RFC5681] requires that
any
> segments sent as part of the Limited Transmit
mechanism not
> be counted in FlightSize for the purpose of the
above
> equation.
>
> IIUC the segments P21..P29 in your example were sent because
of Limited
> Transmit, and so don't count. The flightsize for the purposes
of (4.2) is
> therefore 20 after both losses, and the cwnd does not go up
on the second
> loss.
>
> (2)
> " Even a single shot burst every time there is significant
loss
> event is not acceptable, not to mention continuous
aggressiveness, and
> this is exactly what RFC 2914 and RFC 5033 explicitly address
and warn
> about."
>
> "Significant loss event" is the key phrase here. The intent
of TLP/PTO is to
> equalize the treatment of a small packet loss whether it
happened in the
> middle of a burst or the end. Why should an isolated loss be
treated
> differently based on its position in the burst? This is just
a logical
> extension of fast retransmit, which also modified the RTO
paradigm. The
> working group consensus is that this is a feature, not a bug;
you're welcome
> to feel otherwise but I suspect you're in the rough here.
>
> Regards
> Martin
>
>
> On Wed, Dec 16, 2020 at 4:11 PM Markku Kojo
<kojo@xxxxxxxxxxxxxx> wrote:
> Hi Martin,
>
> See inline.
>
> On Wed, 16 Dec 2020, Martin Duke wrote:
>
> > Hi Markku,
> >
> > There is a ton here, but I'll try to address the top
points.
> Hopefully
> > they obviate the rest.
>
> Sorry for being verbose. I tried to be clear but you
actually
> removed my
> key issues/questions ;)
>
> > 1.
> > [Markku]
> > "Hmm, not sure what you mean by "this is a new loss
detection
> after
> > acknowledgment of new data"?
> > But anyway, RFC 5681 gives the general principle to
reduce
> cwnd and
> > ssthresh twice if a retransmission is lost but IMHO
(and I
> believe many
> > who have designed new loss recovery and CC algorithms
or
> implemented
> > them
> > agree) that it is hard to get things right if only
congestion
> control
> > principles are available and no algorithm."
> >
> > [Martin]
> > So 6675 Sec 5 is quite explicit that there is only
one cwnd
> reduction
> > per fast recovery episode, which ends once new data
has been
> > acknowledged.
>
> To be more precise: fast recovery ends when the current
window
> becomes
> cumulatively acknowledged, that is,
>
> (4.1) RecoveryPoint (= HighData at the beginning)
becomes
> acknowledged
>
> I believe we agree and you meant this although new data
below
> RecoveryPoint may become cumulatively acknowledged
already
> earlier
> during the fast recovery. Reno loss recovery in RFC
5681 ends,
> when
> (any) new data has been acknowledged.
>
> > By definition, if a retransmission is lost it is
because
> > newer data has been acknowledged, so it's a new
recovery
> episode.
>
> Not sure where you have this definition? Newer than
what are you
> referring to?
>
> But, yes, if a retransmission is lost with RFC 6675
algorithm,
> it requires RTO to be detected and definitely starts a
new
> recovery
> episode. That is, a new recovery episode is enforced by
step
> (1.a) of
> NextSeg () which prevents retransmission if a segment
that has
> already
> been retransmitted. If RACK-TLP is used for detecting
loss with
> RFC 6675
> things get different in many ways, because it may
detect loss of
> a
> retransmission. It would pretty much require an entire
redesign
> of the algorith. For example, calculation of pipe does
not
> consider
> segments that have been retransmitted more than once.
>
> > Meanwhile, during the Fast Recovery period the
incoming acks
> implicitly
> > remove data from the network and therefore keep
flightsize
> low.
>
> Incorrect. FlightSize != pipe. Only cumulative acks
remove data
> from
> FlightSize and new data transmitted during fast
recovery inflate
> FlightSize. How FlightSize evolves depends on loss
pattern as I
> said.
> It is also possible that FlightSize is low, it may err
in both
> directions. A simple example can be used as a proof for
the case
> where
> cwnd increases if a loss of retransmission is detected
and
> repaired:
>
> RFC 6675 recovery with RACK-TLP loss detection:
> (contains some inaccuracies because it has not been
defined how
> lost rexmits are calculated into pipe)
>
> cwnd=20; packets P1,...,P20 in flight = current window
of data
> [P1 dropped and rexmit of P1 will also be dropped]
>
> DupAck w/SACK for P2 arrives
> [loss of P1 detected after one RTT from original xmit
of P1]
> [cwnd=ssthresh=10]
> P1 is rexmitted (and it logically starts next window of
data)
>
> DupAcks w/ SACK for original P3..11 arrive
> DupAck w/ SACK for original P12 arrives
> [cwnd-pipe = 10-9 >=1]
> send P21
> DupAck w/SACK for P13 arrives
> send P22
> ...
> DupAck w/SACK for P20 arrives
> send P29
> [FlightSize=29]
>
> (Ack for rexmit of P1 would arrive here unless it got
dropped)
>
> DupAck w/SACK for P21 arrives
> [loss of rexmit P1 detected after one RTT from rexmit
of P1]
>
> SET cwnd = ssthresh = FlightSize/2= 29/2 = 14,5
>
> CWND INCREASES when it should be at most 5 after
halving it
> twice!!!
>
> > We can continue to go around on our interpretation of
these
> documents,
> > but fundamentally if there is ambiguity in 5681/6675
we should
> bis
> > those RFCs rather than expand the scope of RACK.
>
> As I said earlier, I am not opposing bis, though
5681bis wuold
> not
> be needed, I think.
>
> But let me repeat: if we publish RACK-TLP now without
necessary
> warnings
> or with a correct congesion control algorithm someone
will try
> to
> implement RACK-TLP with RFC 6675 and it will be a total
mesh.
> The
> behavior will be unpredictable and quite likely unsafe
> congestion
> control behavior.
>
> > 2.
> > [Markku]
> > " In short:
> > When with a non-RACK-TLP implementation timer (RTO)
expires:
> cwnd=1
> > MSS,
> > and slow start is entered.
> > When with a RACK_TLP implementation timer (PTO)
expires,
> > normal fast recovery is entered (unless implementing
> > also PRR). So no RTO recovery as explicitly stated in
Sec.
> 7.4.1."
> >
> > [Martin]
> > There may be a misunderstanding here. PTO is not the
same as
> RTO, and
> > both mechanisms exist! The loss response to a PTO is
to send a
> probe;
> > the RTO response is as with conventional TCP. In
Section 7.3:
>
> No, I don't think I misunderstood. If you call timeout
with
> another name, it is still timeout. And congestion
control does
> not
> consider which segments to send (SND.UNA vs. probe w/
higher
> sequence
> number), only how much is sent.
>
> You ignored my major point where I decoupled congestion
control
> from loss
> detection and loss recovery and compared RFC 5681
behavior to
> RACK-TLP
> behavior in exactly the same scenario where an entire
flight is
> lost and
> timer expires.
>
> Please comment why congestion control behavior is
allowed to be
> radically
> different in these two implementations?
>
> RFC 5681 & RFC 6298 timeout:
>
> RTO=SRTT+4*RTTVAR (RTO used for arming the
timer)
> 1. RTO timer expires
> 2. cwnd=1 MSS; ssthresh=FlightSize/2; rexmit one
segment
> 3. Ack of rexmit sent in step 2 arrives
> 4. cwnd = cwnd+1 MSS; send two segments
> ...
>
> RACK-TLP timeout:
>
> PTO=min(2*SRTT,RTO) (PTO used for arming the
timer)
> 1. PTO times expires
> 2. (cwnd=1 MSS); (re)xmit one segment
> 3. Ack of (re)xmit sent in srep 2 arrives
> 4. cwnd = ssthresh = FlightSize/2; send N=cwnd
segments
>
> If FlightSize is 100 segments when timer expires,
congestion
> control is
> the same in steps 1-3, but in step 4 the standard
congestion
> control
> allows transmitting 2 segments, while RACK-TLP would
allow
> blasting 50 segments.
>
> > After attempting to send a loss probe, regardless of
whether a
> loss
> > probe was sent, the sender MUST re-arm the RTO
timer, not
> the PTO
> > timer, if FlightSize is not zero. This ensures
RTO
> recovery remains
> > the last resort if TLP fails.
> > "
>
> This does not prevent the above RACK-TLP behavior from
getting
> realized.
>
> > So a pure RTO response exists in the case of
persistent
> congestion that
> > causes losses of probes or their ACKs.
>
> Yes, RTO response exists BUT only after RACK-TLP at
least once
> blasts the
> network. It may well be that with smaller windows
RACK-TLP is
> successful
> during its TLP initiated overly aggressive "fast
recovery" and
> never
> enters RTO recovery because it may detect and repair
also loss
> of
> rexmits. That is, it continues at too high rate even if
lost
> rexmits
> indicate that congestion persists in successive windows
of data.
> And
> worse, it is successful because it pushes away other
compatible
> TCP
> flows by being too aggressive and unfair.
>
> Even a single shot burst every time there is
significant loss
> event is not acceptable, not to mention continuous
> aggressiveness, and
> this is exactly what RFC 2914 and RFC 5033 explicitly
address
> and warn
> about.
>
> Are we ignoring these BCPs that have IETF consensus?
>
> And the other important question I'd like to have an
answer:
>
> What is the justification to modify standard TCP
congestion
> control to
> use fast recovery instead of slow start for a case
where timeout
> is
> needed to detect the packet losses because there is no
feedback
> and ack
> clock is lost? RACK-TLP explicitly instructs to do so
in Sec.
> 7.4.1.
>
> As I noted: based on what is written in the draft it
does not
> intend to
> change congestion control but effectively it does.
>
> /Markku
>
> > Martin
> >
> >
> > On Wed, Dec 16, 2020 at 11:39 AM Markku Kojo
> <kojo@xxxxxxxxxxxxxx>
> > wrote:
> > Hi Martin,
> >
> > On Tue, 15 Dec 2020, Martin Duke wrote:
> >
> > > Hi Markku,
> > >
> > > Thanks for the comments. The authors will
incorporate
> > many of your
> > > suggestions after the IESG review.
> > >
> > > There's one thing I don't understand in your
comments:
> > >
> > > " That is,
> > > where can an implementer find advice for
correct
> > congestion control
> > > actions with RACK-TLP, when:
> > >
> > > (1) a loss of rexmitted segment is detected
> > > (2) an entire flight of data gets dropped
(and
> detected),
> > > that is, when there is no feedback
available and
> a
> > timeout
> > > is needed to detect the loss "
> > >
> > > Section 9.3 is the discussion about CC, and
is clear
> that
> > the
> > > implementer should use either 5681 or 6937.
> >
> > Just a cite nit: RFC 5681 provides basic CC
concepts and
> > some useful CC
> > guidelines but given that RACK-TLP MUST
implement SACK
> the
> > algorithm in
> > RFC 5681 is not that useful and an implementer
quite
> likely
> > follows
> > mainly the algorithm in RFC 6675 (and not RFC
6937 at
> all
> > if not
> > implementing PRR).
> > And RFC 6675 is not mentioned in Sec 9.3,
though it is
> > listed in the
> > Sec. 4 (Requirements).
> >
> > > You went through the 6937 case in detail.
> >
> > Yes, but without correct CC actions.
> >
> > > If 5681, it's pretty clear to me that in (1)
this is a
> > new loss
> > > detection after acknowledgment of new data,
and
> therefore
> > requires a
> > > second halving of cwnd.
> >
> > Hmm, not sure what you mean by "this is a new
loss
> > detection after
> > acknowledgment of new data"?
> > But anyway, RFC 5681 gives the general
principle to
> reduce
> > cwnd and
> > ssthresh twice if a retransmission is lost but
IMHO (and
> I
> > believe many
> > who have designed new loss recovery and CC
algorithms or
> > implemented them
> > agree) that it is hard to get things right if
only
> > congestion control
> > principles are available and no algorithm.
> > That's why ALL mechanisms that we have include
a quite
> > detailed algorithm
> > with all necessary variables and actions for
loss
> recovery
> > and/or CC
> > purposes (and often also pseudocode). Like this
document
> > does for loss
> > detection.
> >
> > So the problem is that we do not have a
detailed enough
> > algorithm or
> > rule that tells exactly what to do when a loss
of rexmit
> is
> > detected.
> > Even worse, the algorithms in RFC 5681 and RFC
6675
> refer
> > to
> > equation (4) of RFC 5681 to reduce ssthresh and
cwnd
> when a
> > loss
> > requiring a congestion control action is
detected:
> >
> > (cwnd =) ssthresh = FlightSize / 2)
> >
> > And RFC 5681 gives a warning not to halve cwnd
in the
> > equation but
> > FlightSize.
> >
> > That is, this equation is what an implementer
> intuitively
> > would use
> > when reading the relevant RFCs but it gives a
wrong
> result
> > for
> > outstanding data when in fast recovery (when
the sender
> is
> > in
> > congestion avoidance and the equation (4) is
used to
> halve
> > cwnd, it
> > gives a correct result).
> > More precisely, during fast recovery FlightSize
is
> inflated
> > when new
> > data is sent and reduced when segments are
cumulatively
> > Acked.
> > What the outcome is depends on the loss
pattern. In the
> > worst case,
> > FlightSize is signficantly larger than in the
beginning
> of
> > the fast
> > recovery when FlightSize was (correctly) used
to
> determine
> > the halved
> > value for cwnd and ssthresh, i.e., equation (4)
may
> result
> > in
> > *increasing* cwnd upon detecting a loss of a
rexmitted
> > segment, instead
> > of further halving it.
> >
> > A clever implementer might have no problem to
have it
> right
> > with some
> > thinking but I am afraid that there will be
incorrect
> > implementations
> > with what is currently specified. Not all
implementers
> have
> > spent
> > signicicant fraction of their career in solving
TCP
> > peculiarities.
> >
> > > For (2), the RTO timer is still operative so
> > > the RTO recovery rules would still follow.
> >
> > In short:
> > When with a non-RACK-TLP implementation timer
(RTO)
> > expires: cwnd=1 MSS,
> > and slow start is entered.
> > When with a RACK_TLP implementation timer (PTO)
expires,
> > normal fast recovery is entered (unless
implementing
> > also PRR). So no RTO recovery as explicitly
stated in
> Sec.
> > 7.4.1.
> >
> > This means that this document explicitly
modifies
> standard
> > TCP congestion
> > control when there are no acks coming and the
> > retransmission timer
> > expires
> >
> > from: RTO=SRTT+4*RTTVAR (RTO used for arming
the timer)
> > 1. RTO timer expires
> > 2. cwnd=1 MSS; ssthresh=FlightSize/2;
rexmit one
> > segment
> > 3. Ack of rexmit sent in step 2 arrives
> > 4. cwnd = cwnd+1 MSS; send two segments
> > ...
> >
> > to: PTO=min(2*SRTT,RTO) (PRO used for arming
the
> timer)
> > 1. PTO times expires
> > 2. (cwnd=1 MSS); (re)xmit one segment
> > 3. Ack of (re)xmit sent in srep 2
arrives
> > 4. cwnd = ssthresh = FlightSize/2; send
N=cwnd
> > segments
> >
> > For example, if FlightSize is 100 segments when
timer
> > expires,
> > congestion control is the same in steps 1-3,
but in step
> 4
> > the
> > current standard congestion control allows
transmitting
> 2
> > segments,
> > while RACK-TLP would allow blasting 50
segments.
> >
> > Question is: what is the justification to
modify
> standard
> > TCP
> > congestion control to use fast recovery instead
of slow
> > start for a
> > case where timeout is needed to detect loss
because
> there
> > is no
> > feedback and ack clock is lost? The draft does
not give
> any
> > justification. This clearly is in conflict with
items
> (0)
> > and (1)
> > in BCP 133 (RFC 5033).
> >
> > Furthermore, there is no implementation nor
experimental
> > experience
> > evaluating this change. The implementation with
> > experimental experience
> > uses PRR (RFC 6937) which is an Experimental
> specification
> > including a
> > novel "trick" that directs PRR fast recovery to
> effectively
> > use slow
> > start in this case at hand.
> >
> >
> > > In other words, I am not seeing a case that
requires
> new
> > congestion
> > > control concepts except as discussed in 9.3.
> >
> > See above. The change in standard congestion
control for
> > (2).
> > The draft intends not to change congestion
control but
> > effectively it
> > does without any operational evidence.
> >
> > What's also is missing and would be very
useful:
> >
> > - For (1), a hint for an implementer saying
that because
> > RACK-TLP is
> > able to detect a loss of a rexmit unlike any
other
> loss
> > detection
> > algorithm, the sender MUST react twice to
congestion
> > (and cite
> > RFC 5681). And cite a document where
necessary
> correct
> > actions
> > are described.
> >
> > - For (1), advise that an implementer needs to
keep
> track
> > when it
> > detects a loss of a retransmitted segment.
Current
> > algorithms
> > in the draft detect a loss of retransmitted
segment
> > exactly in
> > the same way as loss of any other segment.
There
> seems
> > to be
> > nothing to track when a retransmission of a
> > retransmitted segment
> > takes place. Therefore, the algorithms
should have
> > additional
> > actions to correctly track when such a loss
is
> detected.
> >
> > - For (1), discussion on how many times a loss
of a
> > retransmission
> > of the same segment may occur and be
detected. Seems
> > that it
> > may be possible to drop a rexmitted segment
more than
> > once and
> > detect it also several times? What are the
> > implications?
> >
> > - If previous is possible, then the algorithm
possibly
> also
> > may detect a loss of a new segment that was
sent
> during
> > fast
> > recovery? This is also loss in two
successive windows
> of
> > data,
> > and cwnd MUST be lowered twice. This
discussion and
> > necessary
> > actions to track it are missing, if such
scenario is
> > possible.
> >
> > > What am I missing?
> >
> > Hope the above helps.
> >
> > /Markku
> >
> >
> > <snipping the rest>
> >
> >
>
>
>
