Minutes for ICCRG at IETF-95
minutes-95-iccrg-2

ICCRG meeting, IETF 95, Buenos Aires
Monday April 4, 14:00-15:30, Buen Ayre A
Meeting minutes

Presentation:
David Hayes: Congestion control in recursive network architectures.

Bob Briscoe: When I looked, I did not see layered congestion protocols in the
Internet. Is this real? David: There are proxies. Bob: ... this seems like not
quite the same thing. Jana Iyengar: Chaining is different to just chaining TCP
- the signals from deep in the network needs to control the original source.
David: If it is a controlled source, rather than admission control. Jana: There
could still be interactions if they operate on similar timescales. Coupling
between loops is important. David: Yes, one thing I didn't mention though was
that lower layers perform congestion control on aggregates. Jana: The
interaction with chains is an important area.

Presentation:
Alejandro Popovsky: Peeking at the bottleneck, opportunities for bufferbloat
prevention.

Stuart Cheshire: The upstream direction appears source-limited?
Alejandro: Yes, there were mostly ACKs going upstream. These are mainly web
requests. Stuart: I see. Jana: What is RA on the slide? Alejandro: Connection
rate (bytes/RTT). CA is the number of bytes in flight. Randell Jesup: How would
the advantages you spoke off relate to using HTTP/2? Alejandro: There is gain
when using HTTP/2.

Presentation:
Michael Welzl: TCP in UDP (draft-welzl-irtf-iccrg-tcp-in-udp-00)

Michael: Is the flow label in v6 actually used for ECMP?
Gorry Fairhurst: The current UDP guidelines for ECMP support is to put entropy
in the port space and also now to insert this in the flow info for IPv6. I’d
expect (hope) that routers would use the flow info, it’s likely some systems
will use this already, some will use port information. You could get a mixture.
Gorry: Did you look at proxies and TCP-splitters in the network when
Happy-eyeballing? Michael: Not looked at this. Michael Tuexen: How do you
ensure the right sequence of UDP packets and TCP-in-UDP packets on the server
side, when using multicore systems? Michael: In our cases this worked, but we
didn't have a multicore system. Brian Tramell: Announcement: presentation in
MAPRG, can the Internet work over UDP? Jana: Can you share loss detection
between short flows? Think SCTP and streams... Michael: Not done at the moment.
Jana: There is space here to also do that by using data from other flows in
Fast Recovery, which is good for short flows for TLP. Jana: Bottleneck sharing
could be done today... without the UDP encaps. Matt Sargent: Does this have
parallels with multi-path TCP. Michael: It has some commonalities but also
differences. [ Note: this was later clarified in an email:
http://www.ietf.org/mail-archive/web/iccrg/current/msg01314.html ] -: I don’t
see the point if they do not share the same path. Michael: The UDP
encapsulation forces packets on the same path.

Presentation:
Koen De Schepper: PI^2: PI for Classic and Scalable TCP.
- Questions taken to the list.

Discussion:
What are the ICCRG's roles and how can it be more effective?
- Taken to the list.

Ended 15:34.