Transport protocols and bulk file transfer

[Scott Brim <>]

A lot went on, and I brought home some work:

* Lots of discussion about PIE (Cisco) and CoDel (Jacobsen/Nichols),
  particularly with Fair Queuing.  Both of these are tremendously
  better than what's in endpoints today, and what they are arguing
  about is how to make them even more tremendously better.
  sfq-CoDel (stochastic fair queueing) is production in Linux
  already -- we could deploy it in endpoints tomorrow.  CoDel is
  better across the wide area, PIE is better in data centers.  Both
  still need work but they are already excellent.

* SCTP-NAT is done.  SCTP/UDP is done but what's missing is a Happy
  Eyeballs like decision making mechanism, to decide between
  SCTP-over-UDP versus native SCTP.  Also I'm concerned about
  layering without communication: running happy eyeballs functions
  simultaneously for TCP/SCTP and SCTP[/UDP], in a TCP/SCTP shim
  (very likely), in the kernel, and in the web browser?  To be
  sorted out.  Dan Wing is going to try.

* The RTCWEB people chose to do SCTP over DTLS, not over UDP, for
  security.  The problem is that that pushes the happy eyeballs
  decision up into user space (UDP is in the kernel).  The claim in
  RTCWEB is that DTLS will be in the kernel "someday" so it's okay.
  We'll see.

* David Black (EMC): DSCP markings: there will ONLY be 4 defined
  classes between carriers and all carriers will use them.
  "guaranteed" (EF), "bulk inelastic" (AF41 - low loss, low delay,
  low jitter at high bandwidth - traffic load must be controlled,
  will avoid drops by AQM, may get bursty packet loss), "assured"
  (AF31, AF32, AF33 - to transport traffic without bandwidth
  requirements); and "default"/best effort.

* There's an idea of signaling an end-to-end SLA at the granularity
  of the destination AS.  I don't think this is very useful.

* Gorry Fairhurst: draft-ietf-tcpm-newcwv-00: rate limiting built
  into TCP, instead of it pushing as hard as it can.  Concerns about
  how bursty it's going to be.  Still twiddling pipeack estimation
  and smoothing filters.  Implementation will be ready in next few
  months.

* There's now an Active Queue Management mailing list, since we're
  trying to get organized about it.


* Gorry Fairhurst (U of Aberdeen) - follow up on building rate
  limiting into TCP.  Do some experiments.  This fits well with rate
  limiting TCP (using tc and a hierarchical token bucket).

* Jim Martin (Clemson): 3 topics: campus mobility services; Internet2
  and an experimental R&E MVNO; and transport (he has consulted with
  the cable industry on QoS issues for a long time).  Will follow up
  about possible collaboration.

* Bob Braden and Jim Gettys are compiling a "lessons learned" about
  buffer bloat.  Case studies.  How to do better.  I told him about
  our problem with dropped packets.

* Benno Overeinder (NLNet labs, University of Amsterdam) was part of
  the MPTCP-over-OpenFlow project, and he agrees that doing it as a
  decoupled overlay wasn't so interesting, so he's working on better
  integration between the layers, letting MPTCP have insight into and
  control of OF.  He wants to follow up with us.  He will send some
  research info.

* Benno will also send a paper on transport layer issues.

* Suresh Krishnan (Ericsson) says he's working on experiments with
  OpenFlow and dynamically changing QoS.