TSVWG Virtual Interim (2/20/2020 - 9am-11am US/Eastern)

Posted agenda:
1. Agenda bashing (chairs - 1 min)
2. Plan for Vancouver (David Black's charts - ~10min)
3. Some Congestion Experienced (Morton/Heist/Grimes/etc.? - ~30 min)
4. Update on status and near-term plans for the set of L4S drafts (Bob? - 4 min)
5. Update on L4S & TCP Prague implementation, test, evaluation
(Greg/Bob/Koen? - 15 min)
6. L4S Issues List Discussion (goal is to see if we can agree on what more
needs to be done on each issue to suffice for Experimental/Informational RFCs):
    6.1 Discuss status on (Bob - ~30 min):
          #16 on classic ECN interaction
          #21 CE codepoint semantics (closely related to #16)
          #20 ECT(1) codepoint usage
    6.2  Discuss plans forward to address (Bob - ~15 min):
          #26 on admission control
          #27 on terminology
          #22 on deployment feasibility
          #24 on evaluation & testing results
          #28 on DualQ suitability
    6.3 (We will try to handle the issues below by mailing list prior to the
    meeting)
        If needed, discuss close-out of (~15 min):
          #25 on IPR
          #18 on loss detection in time-units
          #23 on implementation status
          #19 on the single codepoint
          #17 on FQ interaction


1. Agenda bashing - David Black indicated we'd like to make sure to have 30 minutes for SCE discussion.


2. Plan for Vancouver - David Black led discussion "ECT(1) and the Internet"

- TSVWG chairs propose to make decision in Vancouver on what ECT(1) is used for.
- Need to determine for the Internet whether ECT(1) is an input (as in L4s) or output (as in SCE)
- Koen and asked about the distinction between choosing between input/output and L4S/SCE.
- Steve Blake mentioned that L4S could be made compatible with each.
- Use of DSCP to scope this was discussed.  At least with L4S, this is not the current approach (mentioned by Gorry and David).
- Jake Holland asked about the decision not to use DSCP.  David replied that due to DSCP bleaching, there can be problems at AS boundaries.
- Steve Blake noted anyone running these experiments will tinker with boxes, so may handle the DSCP bleaching concern.
- Jonathan Morton said this looks good.
- Jake Holland noted the FQ statement "no coexistence problems" is too strong.  David will adjust this.
- Steve explained that routers will have to change for either approach.  Gorry and David answered there may be some routers that haven't changed on the path.
- Question of "what is starvation" should be taken offline.


3. Some Congestion Experienced - presented by Jonathan Morton and Pete Heist

- This covered 4 drafts (sce, codel-approx-fair, lightweight-fair-queueing, cheap-nasty-queuing).
- Asking for SCE adoption by TSVWG, but not expecting decision today.
- SCE problem statement and goals summarized.
- Results from several test scenarios were shared, showing how SCE works.
- LFQ was explained and some results with LFQ were shared.
- CNQ-CodelAF (Controlled Delay Approximate Fairness) was introduced and test scenario results shown.
- Approach on RTT fairness will be to adapt CUBIC.

- Greg White asked about LFQ and uses of FIFOs.  Jonathan explained 3 FIFO construction for recirculation.
- Greg asked if AF eliminated the latency benefit for SCE.  Jonathan explained this is due to a single queue.
- The point of CNQ-CodelAF is that SCE can use the spare capacity.

- Gorry asked about latency benefits.  Jonathan explained eliminating the sawtooth.
- David clarified question as what needs to be done in the network to acheive lower latency for SCE flows.
- Jonathan explained best option is FQ, and will discuss more w/ Gorry offline.
- Pete explained CodelAF in single queue is an approach to incremental deployment.
- Bob mentioned it should work right away in order to get deployed, or there is no benefit to deploying.
- Pete pointed out improved link utilization.
- Greg White asked if data is anecdotal cases and time series, rather than wider range, CDFs.


4. L4S Draft Status - updates discussed by Bob Briscoe

- overviewed changes to drafts, some just recently uploaded.


5. L4S and TCP Prague status - discussed by Bob Briscoe

- L4S implementation status overview provided.
- Only reference implementation discussed here, others for IETF 107.
- Evaluation in progress for reducing RTT dependence.
- classic bottleneck detection algorithm discussed (for issue #16)
  - score-based gradual transition, not modal
  - should take question offline on hysteresis and oscillation; Bob's answer is around stickiness rather than hysteresis.
  - paper available is just the design, evaluation is starting, Linux TCP Prague work by Asad Ahmed
- RTT Independence in TCP Prague work by Olivier Tilmans & Koen De Schepper
  - described impact of queuing delays and shorter L4S queues
  - explained new Prague add-on to steer RTT dependence, controlling additive increase and EWMA update frequency
- demo video played
  - Sebastian Moeller asked about the issue with dual queue's only approximate sharing, and short L4S queue crowding out the classic queue
  - Koen responded about policing functionality in the network at appropriate places.
  - There should be continued discussion on the distinction between requirements for flows to get into the L4S queue versus means to enforce them.
- Accurate ECN feedback in TCP implemented by Ilpo Jaarvinen in Linux
- Paced chirping code is being worked on in TCP Prague (Joakim Misund)
- Bob says all code has been available for some time.


6. L4S Issues List Discussion
- Issue #16 - interaction w/ 3168-only ECN
  - should still remain open, but making progress on detection/fallback

Meeting officially concluded based on time; some participants continued on the webex session to discuss additional issues list items as "hallway conversation".