[Rdma-cc-interest] Minutes from the IETF-113 side meeting on 802.1 Congestion Management Initiatives

[Paul Congdon <[e-mail]>]

Here are the minutes from the side meeting on March 23, 2022 - at 10AM in
Vienna.  Please send corrections or comments to the mailing list.

Thanks for your participation,
Paul

1.       Attendees:

a.       Paul Congdon, Tallac Networks, Huawei

b.       Richard Scheffenegger, NetApp

c.       Hirochika Asai, Preferred Networks / WIDE Project

d.       Stein Gjessing, University of Oslo

e.       Michael Welzl, University of Oslo

f.        José Duato  (online)

g.       V. Michael Boyle (GOV) (online)

h.       Lily  Lv (来宾) (Guest) (online)

i.         Tianran (来宾) (Guest) (online)

j.         Haibo Wang (来宾) (Guest) (online)

k.       huafeng wen (来宾) (Guest) (online)

l.         Dmitry Afanasiev (Guest) (online)

m.     Marco Liebsch (NEC) (Guest) (online)


2.       Paul Congdon presented the background slide material.  The slides
may be obtained from:
https://www.ieee802.org/1/files/public/docs2022/new-congdon-IETF-113-sidemeeting-0322-v01.pdf
.


3.       Jose Duato noted that calculating headroom on a per-traffic class
basis must be done assuming the worst case, that is, packets are in flight
for only a single traffic class at a time.  In reality, packets on the wire
may be from multiple traffic classes and thus we could be overestimating
the needed headroom.  This is traffic pattern dependent and can’t be
predicted, so a worst case needs to be assumed.  Jose mentioned some
implementations could have a shared buffer for ingress that acts like a
“landing pad” for packets before being marshalled out to the intended
traffic class.  Only this ‘landing pad’ would require implementing
headroom, independent of traffic class.


4.       Richard Scheffenegger pointed out that the traditional end-to-end
congestion control loop is too slow for these types of data center networks.


5.       Richard voiced concerns that SFC may have some challenges during
standardization.  There may be hidden complexities.  Jose spoke in favor of
SFC speeding up backpressure (i.e., backward propagation of flow control)
and the fact that individual flow information is contained in the signaling
message could be helpful to stop only the congesting flows, thus avoiding
head of line blocking.  Jose voiced that SFC is a good idea, but it’s
better to be combined with accurate congestion detection mechanism.
Inaccurate congestion detection may worse the performance. Prof. Duato is
doing research on congestion detection mechanisms and believes this can
help.


6.        Jose noted that Clos network topology may not be the most
cost-effective design as the network scales in size.  A consideration of
Dragonfly topologies may be more cost-effective for actual datacenter
traffic, but could involve other changes to deadlock avoidance and
congestion management.


7.       Hirochika mentioned that issues associated to overlay networks and
SFC should be discussed in the IETF.


8.       Michael Welzl voiced a concern about reactive congestion control
algorithms in general.  He wondered if in-band telemetry schemes, such as
HPCC+, are practical to implement and for deployments – perhaps their
implementation cost on NICs could be problematic?  He also asked if
‘shadow’ queue approaches are practical and more cost effective to
implement?  Michael references a paper from Nick McKeown related to
congestion control that is relevant. See:
http://yuba.stanford.edu/~nickm/papers/p230-cronkite-ratcliff.pdf

Meeting adjourned at 11:35AM