Re: [tcpm] Questions about TLP

[Yuchung Cheng <[e-mail]>]

Sorry for the late reply. Forget to hit send after drafting the reply...

Thanks for spotting that. TCP_send_probe() should always re-arm the
RTO regardless of a TLP is sent or not, if there's unacknowledged
data. Proposed rev in

"Note that the sender MUST arm an RTO timer and not the PTO timer *at the
end of TLP_send_probe() as long as FlightSize is not zero, regardless of
sending a probe or not*. This ensures that the sender does not send
repeated, back-to-back TLP probes. *Also checking TLPRxtOut prior to
sending the loss probe* is important to avoid TLP loops if an application
writes periodically at an interval less than PTO."

*From: *Yi Huang <huanyi=[e-mail]>
*Date: *Thu, May 2, 2019 at 11:47 AM
*To: *Yuchung Cheng, Matt Olson
*Cc: *[e-mail]

If TLPRxtOut is checked in TLP_send_probe(), a new PTO can cancel an
> already running RTO timer armed by sending TLP previously. If
> TLP_send_probe() decides not to send the probe for the new PTO, what should
> the sender do next? Rearm RTO or treat it as RTO timeout? The draft states
> we must arm an RTO after transmitting TLP but in this case, no probe will
> be sent.
>
> Also, in page 17, the draft says "This is important to avoid TLP loops if
> an application writes periodically at an interval less than PTO." but if
> the app writes periodically at an interval less than PTO, PTO will just be
> pushed out and not fire until the sender cannot send anything due to window
> limit. Then in this case, TLP loops do not seem to exist if TLP loops means
> back-to-back TLP probes. Am I missing anything here?
>
> Thanks,
>
> Yi
>
> -----Original Message-----
> From: Yuchung Cheng <ycheng=[e-mail]>
> Sent: Thursday, May 2, 2019 10:25 AM
> To: Matt Olson <[e-mail]>
> Cc: Yi Huang <[e-mail]>; [e-mail]
> Subject: Re: [tcpm] Questions about TLP
>
> On Thu, May 2, 2019 at 10:09 AM Matt Olson <[e-mail]> wrote:
> >
> > Also, section 6.5.1 Step 1 is presented as a complete set of conditions
> for scheduling a PTO, but it is missing a bullet point for TLPRxtOut.
> Thanks for checking: but checking TLPRxtOut is not necessary and is not
> preferred when scheduling a probe -- the goal is to prevent more than one
> probe inflight, so checking right before TLPRxtOut sending the next one
> allows the best coverage of application-limited writes
> (burst-idle-burst-....).
>
>
>
>
> >
> > -----Original Message-----
> > From: Yuchung Cheng <[e-mail]>
> > Sent: Wednesday, May 1, 2019 11:33 PM
> > To: Yi Huang <huanyi=[e-mail]>
> > Cc: [e-mail]; Matt Olson <[e-mail]>
> > Subject: Re: [tcpm] Questions about TLP
> >
> > On Wed, May 1, 2019 at 10:48 PM Yi Huang <huanyi=
> [e-mail]> wrote:
> > >
> > > Hi RACK/TLP authors,
> > >
> > >
> > >
> > > I have the following questions (page numbers refer to
> draft-ietf-tcpm-rack-05):
> > >
> > >
> > >
> > > 1.In page 16, it says “Finally, if the time at which an RTO would fire
> (here denoted "TCP_RTO_expire") is sooner than the computed time for the
> PTO, then a probe is scheduled to be sent at that earlier time.” What does
> this TCP_RTO_expire actually mean? Does it mean there is another RTO timer
> (possibly started some time in the past) running along with PTO or just RTT
> + 4*RTTVar + Now()? Also, why would another probe be sent at RTO expiration
> time instead of treating it like RTO and collapsing the cwnd?
> >
> > TCP_RTO_expire() should return a timeout value for regular RTO, i.e.
> > SRTT + 4*RTTVAR + Now()
> > We use TCP_RTO_expire() because many implementations including Linux
> > do not use the exact RFC formula
> >
> > The reason it's not treated as a regular RTO is to avoid resetting
> congestion window to 1. The rationale is if the probe was sent within
> min(PTO, RTO) and was delivered successfully, there's no need to reset
> congestion window and re-start slow-start, similar to the rationale of
> Reno's fast recovery reducing window to ssthresh instead of 1. We have
> found this strategy benefits wireless connections, as the chance of
> spurious RTO is high due to the delay variation.
> >
> > >
> > >
> > >
> > > 2.In page 18 section 6.6, a new variable TLPRxtOut is defined and it
> is stated that TLPRxtOut is used to guarantee that there is only one
> outstanding TLP retransmission.. However, it is not clear to me when
> TLPRxtOut should be really used. Should it be used in 6.5.1 Step. 1
> (checking whether we should and can schedule a TLP) or in TLP_send_probe()?
> >
> > This is in
> >
> > 6.6.2.  Recording loss probe states
> >
> >    Senders MUST only send a TLP loss probe retransmission if TLPRxtOut
> >    is false.  This ensures that at any given time a connection has at
> >    most one outstanding TLP retransmission.  This allows the sender to
> >
> > but I agree it'd be more clear to include this in TLP_send_probe()
> pseudo code. I'd incorporate in the next rev.
> > >
> > >
> > >
> > > Thanks,
> > >
> > >
> > >
> > > Yi
> > >
> > >
> > >
> > > _______________________________________________
> > > tcpm mailing list
> > > [e-mail]
> > > https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww
> ..
> > > ietf.org%2Fmailman%2Flistinfo%2Ftcpm&amp;data=01%7C01%7Cmaolson%40mi
> > > cr
> > > osoft.com%7C66db44672e3d4101d72908d6cec82001%7C72f988bf86f141af91ab2
> > > d7
> > > cd011db47%7C1&amp;sdata=n8rp5Emowm459iSKolhIP1hFXBs8pZjsGG2iy3OBAno%
> > > 3D
> > > &amp;reserved=0
> _______________________________________________
> tcpm mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tcpm
>