[tcpm] Tail loss probe talk [was Re: Draft agenda for IETF 87 in Berlin]

[Yuchung Cheng <[e-mail]>]

On Wed, Jul 17, 2013 at 2:01 AM, Scharf, Michael (Michael)
<[e-mail]> wrote:
> Hi all,
>
> The first draft of the TCPM agenda can be found at https://datatracker.ietf.org/meeting/87/agenda/tcpm/ and is also copied below.
>
> We tried to take into account all requests, but as a result, the agenda is pretty full. This implies:
>
> - The slots are *total* time. Authors, please reserve 50% of your time for questions and discussion.
>
> - We got time slot requests for drafts that were not be updated since the last meeting, typically promising new results for the presentation in the meeting. Mailing list discussion in the next week would be *highly* welcome in order to justify a time slot on our full agenda.
Sure we should do that for TLP

>
> Please let us know if you have any comments, suggestions, or if we missed something.
>
> Happy agenda bashing!
>
> Michael, Pasi, Yoshifumi
>
>
>
> *************************************************************************************
>
> TCPM Agenda
> IETF 87 in Berlin, Germany
> Tuesday, July 30, 2013, 09:00-11:30
> ***********************************
>
> WG status
> ---------
>
> 09:00
> TCPM status
> Chairs
> 5min
>
>
> WG items
> --------
>
> 09:05
> draft-ietf-tcpm-1323bis-14
> Richard Scheffenegger
> 20min
>
> 09:25
> draft-ietf-tcpm-fastopen-04
> Jerry Chu / Yuchung Cheng
> 15mins
>
> 09:40
> draft-ietf-tcpm-newcwv-02
> Rafaello Secchi / Gorry Fairhurst
> 20min
>
> 10:00
> draft-ietf-tcpm-rtorestart-00 (no new draft)
> Anna Brunstrom
> 10min
>
> 10:10
> draft-ietf-tcpm-accecn-reqs-02
> Mirja Kuelewind
> 5min
>
>
> Non-WG items
> ------------
>
> 10:15
> draft-zimmermann-tcpm-tcp-rfc4614bis-02
> Alexander Zimmermann
> 10min
>
> 10:25
> draft-dukkipati-tcpm-tcp-loss-probe-01 (no new draft)
> Yuchung Cheng
> 10min
TCP Loss probe (TLP) is not updated the draft because we weren't sure
(yet) what to update, and the new data shows similar improvement that
we've presented
http://static.googleusercontent.com/external_content/untrusted_dlcp/research.google.com/en//pubs/archive/41217.pdf

Based on the recent WG and private discussions and the SIGCOMM
reviews, people are not sure the performance improvement justifies
this new feature, as the loss recovery is feature rich and complicated
with RFC3517, RFC6937, RFC5682, RFC5827, etc.

To simplify, I'd like to present TLP as a simple conceptual change to
RTO recovery: On first timeout, retransmit the last packet instead of
the first unacked packet, and don't change but reduce first timeout to
2 SRTT, as a "quick loss probe".

As an alternative we've tried to only crank down the RTO but the
result is not good, because reducing cwnd to 1 is catastrophic, and
TCP today has to live sudden delay spikes due to buffer bloat and
mobile delays. It's better to be very conservative before declaring
the network is busted and use LW=1.

We are hopeful to get more WG interests for TLP :)

>
> 10:35
> draft-flach-tcpm-fec-00
> Tobias Flach
> 15min
>
> 10:50
> draft-gont-tcpm-tcp-seq-validation-00 (no new draft)
> Fernando Gont
> 10min
>
> 11:00
> draft-kuehlewind-tcpm-ecn-fallback-00
> draft-kuehlewind-tcpm-accurate-ecn-02
> Mirja Kuehlewind
> 15min
>
>
> If time permits
> ---------------
>
> Wired PRR effects
> Mirja Kuehlewind
> 5min
>
> Dealing with sequence-number randomizing firewalls
> Christoph Paasch
> 5min
>
> packetdrill: Scriptable Network Stack Testing, from Sockets to Packets
> Yuchung Cheng
> 5min
> _______________________________________________
> tcpm mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tcpm