[Rdma-cc-interest] Data Center Congestion Control

[Paul Congdon <[e-mail]>]

Hello RDMA CC list participants,

There were two interesting presentations at IETF-108 related to Data Center
Congestion Control.  A couple of discussion points came up that have
surfaced before, so I thought it might be worth highlighting them for
ongoing discussion.

The two presentations in the second session of TSVWG were:

1. HPCC++: Enhanced High Precision Congestion Control - link
<https://datatracker.ietf.org/meeting/108/materials/slides-108-tsvwg-sessb-81-hpcc-enhanced-high-precision-congestion-control>
2. PFC free low delay control protocol - link
<https://datatracker.ietf.org/meeting/108/materials/slides-108-tsvwg-sessb-82-pfc-free-low-delay-control-protocol>

Both of these presentations discussed improving congestion control for
RoCEv2 (as well as other protocols), which re-surfaced a couple of issues:

1. RoCEv2 is specified by IBTA, so the IETF should not propose 'specific'
changes for this protocol.
2. If work were to be done in the IETF, where should it begin; ICCRG or
TSVWG.

I propose the following comments on these issues:

1. Agreed that RoCEv2 and it's DCQCN are under the stewardship of IBTA.
Perhaps what is really needed from IETF is a 'data center' focused CC
approach that any UDP based protocols can use - in the 'data center'. This
was briefly discussed several IETFs ago in a side meeting.  Maybe some of
the tenants of that protocol could come from the recent presentations and
include:
   - Use of in-band signaling (i.e. IOAM) to provide more precise
congestion information to end-points in order to allow a CC to react
quickly and accurately (not overdoing it or under doing it) - motivated by
HPCC.
   - Specify how ECN thresholds can be fluid/dynamic to allow protocols to
start fast and stabilize quickly without congestion  - motivated by LDCP

2. The leadership of IETF have consistently stated that congestion control
ideas should/could start off in ICCRG and then move to TSVWG when/if a
standards track draft is needed.

Just some thoughts... Would love to hear what others have to say.

Paul