Skip to main content

Minutes for AQM at IETF-93
minutes-93-aqm-2

Meeting Minutes Active Queue Management and Packet Scheduling (aqm) WG
Date and time 2015-07-20 16:50
Title Minutes for AQM at IETF-93
State Active
Other versions plain text
Last updated 2015-07-28

minutes-93-aqm-2
WG items
 --------

 19:00  AQM Characterization Guidelines Update  draft-ietf-aqm-eval-guidelines 
 Nicolas Kuhn et. al.  5 min

Gorry Fairhurst: this doc will never be perfect, it's valuable now, we should
proceed to WGLC

 19:05 draft-ietf-aqm-ecn-benefits Gorry Fairhurst 5 min

Richard Scheffenegger: this is now in WGLC, last chance to speak

 19:10 draft-hoeiland-joergensen-aqm-fq-codel Toke Hoeiland-Joergensen  5 min

Andrew McGregor: codel doc is not ready yet but we shouldn't wait
Bob Briscoe: this is a good doc, still haven't finished review

================

 Algorithm Proposals
 -------------------

 19:15 Questioning the goal of hard Delay Targets Scaling AQM config with load
 and capacity Bob Briscoe 10 min

show of hands: half of the audience understands the draft
Stuart Cheshire: agree that loss is more harmful than we think but ECN should
make it go away Bob Briscoe: agree but we then need a good algorithm for ECN
marking Dave Dolson: what are the units of the axis Bob: see the draft, slides
are simplified Matt Mathis: this is not specific to TCP, any congestion control
needs to tradeoff loss rate vs latency vs data rate
    this is why TCP isn't great for measurements
    in old RED, smoothing function is broken
Bob: we need to work on this
Matt: I suggest 1 way exponential moving average
Stuart: we shouldn't blame TCP for everything, this is inherent to congestion
control ?    : if you don't have ECN, should we still use this? Bob: if you
don't have ECN, maybe go for something better than Droptail and use pie or
(fq)codel Matt: both of these are 100x better than drop tail so still do
something

 19:25 Dual Queue Coupled AQM  Koen de Schepper 20 min

Stuart Cheshire: are these served round robin?
Koen de Schepper: classic queue is only served when L4S queue is empty (strict
queuing) Dave Taht: If you are using only the ECT to differentiate flows, how
does legacy Cubic+ECN compare? Koen: this assumes that ECN implies DCTCP, we
should discuss another way to differentiate but not now Mirja: to clarify,
other congestion control algorithms that are as aggressive could work too Koen:
agree Stuart: Apple turned on ECN on all iOS/Mac OS current developer seeds
(see WWDC video) please test it and spread the word Bob: we need an identifier
to differentiate TCP-ECN vs DCTCP David Schinazi: what's the plan for
incremental rollout? Richard: let's move this discussion to the list

 19:45 draft-lauten-aqm-gsp Wolfram Lautenschlaeger 5 min

Richard: encourage feedback on this algorithm on the list