Minutes for DETNET at IETF-91
minutes-91-detnet-3

* DRAFT NOTES *

DETNET non WG forming BoF in IETF91

BoF Chairs: Erik Nordmark
            Jouni Korhonen


Approximately 55 attendees
Jouni's introductions
Erik's summary

** Norm's presentation
A comment came in that stated the first page of Norm's presentation
already presented a solution. Discussion about lossless Ethernet may
not be a solution because it uses flow control which creates nasty
latency problems.

Greg Mersky: Use collision avoidance techniques to accomplish low packet loss.
???: How is network managed? Single administrator or multiple administrators?
???: How is traffic from AVB nodes protected from non-AVB nodes?  The SRP Domain handles this.
Cisco: What is pre-emption?
Pat Thaler: Stop part way into transmission, then send high-priority, then
    send remainder of low-priority packet.
Pat Thaler explained it in detail (she's the IEEE 802.3br editor).
Shitanshu @ Cisco: Does this need to support a mix of requirements?
    Sometimes yes if all devices must receive it to be acceptable to
	the application requirement.

Bob Briscoe: Don't necessarily need reservations. That is a solution.
Lars: What about corruption loss?
Norm: send multiple times
Pat Thaler: Don't retransmit
David (Mellanoz): Why not use loss-less Ethernet?
Pat: Doesn't get low latency using flow control
Shitanshu: What about Jitter
Norm: depends on application

Red? Collision avoidance?

???: ATM is a reference?
Norm: ATM didn't do best effort well - cost-effectively

Bob: Requirements: multi-tenant?
Norm: single administrator; multi-vendor interop
Norm: sports arena - multiple networks


** Pascal - Scoping the work

...

** Erik - IntServ

NSIS people may have already resolved several of these problems.
what was different wrt DetNet?

Bob Briscoe: multi-domain? Implies re-shaping at boundary
Bob: Doing this in an existing production network.
    But with shaping of best-effort traffic (well, all traffic) at the edge
Bob: Data center - High-bandwidth ultra-low latency (HULL)

** Daniel King - PCE

Pascal Thubert: How can we publish the topology of a network with PCE?  TED
    and Link State database can provide this. Need for abstraction of
	topology?
Norm: mixing layer-2 and layer-3 controllers can help solve this problem,
    but it's kind of ugly because various suppliers may have custom
	controllers that cannot interoperate.
Adrian: Abstract topology information. Being done I2RS
Adrian: Perhaps doing a Yang model for describing topologies.

PCE assumes connections - TE environment


** Pascal Thubert - 6TiSCH

** Shitanshu Shah (Cisco) - Deterministic Phb (with DSCP)

** Discussion - open mike

Lars Eggert: Why are customers looking for layer 3 supports?
Norm: Some customers don’t see any need.  If it can’t be done at layer 2
    why can it be done at layer 3?  Layer 3 needs to be able to take
	advantage of layer 2 hardware support.
Lars: BCPs for how done at L2. Gets harder to move up the stack
Norm: Don't want L3 to get in the way
Adrian: Pseudo-wires is a potential technology
Norm: 2 people have mentioned that

Pseudo wires may be a good solution for redundancy.
We have layer 3 solutions, we have layer 2 solutions, what more do we need?
We need the ability to access the layer 2 solutions from layer 3.

Lars: BCPs for how done at L2. Gets harder to move up the stack
Norm: Don't want L3 to get in the way
Lew Burger: Flow identification?  Is this the problem?   What is it we need
    here? We need an inventory of all possible solutions.  If we believe this
	is already solved it would be good to discover this.  Identifying flows
	is a big question.
Bob: Some solutions scale by aggregates but here we have micro flows
Norm: Potential to aggregate; MPLS could do
Erik: number of micro flows in different use cases
Pascal: 6TiSCH found that when it comes to actual solving this problem from
    top to bottom uncovered a lot of holes in all the individual solutions.
	This WG would try to develop a full top-to-bottom solution that can take
	advantage of layer 2.

There was a request to have Norm and Pascal each create a problem
description and do gap analysis to see what’s missing.  Perhaps
doing this for each possible solution.

Pascal: look at what is missing
Ted: PS + gap analysis for both problems (Norm's and Pascal's)