Minutes for DetNet Interim - 2023/13 (virtual)

WG ICS: https://datatracker.ietf.org/meeting/upcoming.ics?filters=detnet

Datatracker: https://datatracker.ietf.org/group/detnet/about/

Tuesday, December 12, 2023

13:00 - 15:00 (UTC)
https://www.timeanddate.com/worldclock/fixedtime.html?msg=DetNet+Interim&iso=20231212T13&p1=1440&ah=2

Materials:
https://datatracker.ietf.org/meeting/interim-2023-detnet-13/session/detnet

Note taking:
https://notes.ietf.org/notes-ietf-interim-2023-detnet-13-detnet?both
Meetecho:
https://meetings.conf.meetecho.com/interim/?group=b4c60a10-3026-422e-a9f2-65affb467bd0

Chat: https://zulip.ietf.org/#narrow/stream/detnet
Session ICS:
https://datatracker.ietf.org/meeting/interim-2023-detnet-13/sessions/detnet.ics

YouTube: https://www.youtube.com/watch?v=17_g8nzAd0E

Slot Start Time (UTC) Duration Information

#1 13:00 5 min Title: Intro

Presenter: János Farkas

Toerless Eckert: Please share plans with next steps about taxonomy on
the list to avoid surprises.

János Farkas: Sure.

#2 13:05 45 min Title: P802.1DC and P802.1Qdv Overview

Presenter: Norm Finn
(start: 13:10)

P802.1DC Quality of Service Provision by Network Systems

This is a stand-alone document (not an amendment to 802.1Q)

Florian Kauer: Difficulties in configuring priority to traffic class mapping and
flow classification.
Norm Finn: Priority (layer 2, i.e., 802.1p) is a bridge feature, not a router feature.
Flow classification (e.g., based on DSCP) is handled via other router mechanisms.
Ballot comment resolution is expected to clarify this aspect.

Toerless Eckert: Would like to see example of use of 802.1DC with DetNet flow
classification.
Norm Finn: 802.1DC contains example of 802.1CB flow identification.  IETF would
need to do something analogous for DetNet, as DetNet is router-based,
not bridge-based.

Florian Kauer: Have initial implementation of DetNet YANG model with 802.1DC,
will share to list.

Janos Farkas (in chat): We are starting P802.1DC initial SA ballot comment
resolution on the Dec 18 TSN weekly call. Everyone is welcome to join, there
are no prerequisites. See details at:
https://listserv.ieee.org/cgi-bin/wa?A2=STDS-802-1-MINUTES;a010b740.23&S=
or at https://1.ieee802.org/tsn/tsn-tg-meetings/


P802.1Qdv Enhancements to Cyclic Queuing and Forwarding

Toerless Eckert: Is multi-priority new standards material or better explanation
of existing material?
Norm Finn: Original CQF needs queue per bin, example on slide 14 would use 9
on/off queues, which aren't available (max of 8).  Enhancement only needs 3
queues because it can subdivide queues into bins (i.e., multiple bins/queue
allowed, in contrast to original CQF being limited to 1 bin/queue).

Norm Finn: Paper on 802 web site (pointer at end of slide deck) explains how
preemption fits into P802.1Qdv framework.

Jinoo Joung: How is cycle time duration determined?  Has multiple effects.
Norm Finn: A set (repertoire) of cycle times is used, shortest cycle times
may be limited to portions of the network.  Known technology area (e.g., for
POTS telephony).  Shortest cycle time could be less than a frame transmission
time.
Jinoo Joung: Cycle time per flow becomes complex.
Norm Finn: Don't need to do that.  Bins are not per-flow, they are per-priority.

Jinoo Joung: How are bursts handled?
Norm Finn: Burst limiting is at edge, limits each flow's bytes/cycle.
Jinoo Joung: Sending a burst results in latency when edge flattens burst.
Norm Finn: Have to decide how much bandwidth to allocate wrt max bursts.
Allocated bandwidth for bursts is dedicated (use it or lose it).
Jinoo Joung: Understood, more work needed on dealing with bursts.

Toerless Eckert: Comparison to ATS?
Norm Finn: If ATS scheduler driven by large-grain clock, result has some
similarities to this multi-queue CQF.
Norm Finn: For continuous flows, wind up buffering 1 packet/hop, ATS and
CQF do this. ATS has to precompute buffering, that computation becomes
intractable at scale.  CQF is similar, but only optimal for flows that
exactly match the bandwidths selected for CQF.

Toerless Eckert: CQF has to bound clock drift to keep bins roughly aligned.
New one-way bin alignment protocol is less demanding than clock synchronization.

Jinoo Joung: If the duration of a cycle is determined by the number of flows
and their packet sizes, then the highest priority cycle can be very large,
and the lower priority cycle duration is unnecessarily even larger?
Norm Finn: Need sufficient buffering in each bin to accomodate flow
interactions (e.g., if all flows transmit at same time). Per hop delay
at node is inversely proportional to packet rate - allowance for bursts
increases that.  Overall "even larger" conclusion is accurate - there are
limits on what is possible.

TongTong Wang: Priorities are per-queue, not per-flow.
Norm Finn: Agree, priorities are bandwidth classes, to which flows are assigned.

Jinoo Joung: Byte-counting per flow is per-flow state.
Norm Finn: Agree, does not scale in general, should not use everywhere, e.g.,
use only at boundaries of frequency locked regions.  Flow aggregation helps
where it is used (count bytes on aggregated flows instead of individual flows).

Norm Finn: In general, need to combine time-based and packet-label-based mechanisms
for a workable flow aggregation solution - can't solely use one or the other class
of mechanism.

Jinoo Joung: The sucess of Interenet is based on statistical multiplexing; whereas TSN and DetNet seem going back to TDM.
Norm Finn: Agree, TSN (especially CQF) and DetNet  provide reliable TDM.
services piggybacked on non-TDM networking technology, e.g., Ethernet.

Toerless Eckert: Goal is to get reliable control loops without sophisticated math .

Norm Finn: What I presented does not require any sophisticated math.

TongTong Wang: TSN has more than just TDM mechanisms to deliver bounded latency.

Norm Finn: Would like DetNet to consider these two 802 docs as part of scaling-up
work.

#3 15:01 Adjourn