ICCRG Research Group Agenda - IETF 126
- Time and Date
-- Wednesday July 22, 09:00-11:00 CEST
- Location: Park Suite 7
Presentations
Chair slides & Hackathon update - Chairs, 5 minutes
Pacing in Transport Protocols, draft-irtf-iccrg-pacing - Michael Welzl, on-site, 5 minutes
- Michael Tüxen: there's a special number from Apple (burst size of
244 microseconds worth of data). would be cool to know where it
comes from.
- Christian Huitema: trade-off between Prague-related considerations
(avoiding bursts) and the need of some minimum burst size (of 10
packets) for GSO to be effective. could be that Apple has based it
on this.
- Michael Tüxen (proxy for someone else): number is 1/4096.
- Chat:
-- Alan Jowett: 244us is approximately the reciprocal of 4096 so,
4096 times per second == 244us
-- Michael Welzl: Yes - probably: 1) less than 1ms, to avoid the
bursts for L4S as Christian said, 2) why not 1/4, for example; 3)
let's implement 1/4 efficiently in the kernel.
Windowless Cumulative ACK Extension for RDMA Retransmission, draft-chen-rdma-windowless-ack - Danyang Chen, remote, 5 minutes
Testing Congestion Control and Queue Management Mechanisms, Mohit Tahiliani, on-site, 15 minutes
- Simone: what about an evaluation framework for congestion control?
-- Mohit: still working on an evaluation suite that is compliant to
RFC 9743
Using Little’s law queueing theoretic approach as congestion window update, Dumisa Ngwenya, on-site, 20 minutes
- Roland Bless: high throughput, low delay not mutually exclusive. did
you test together with loss-based congestion control as delay-based
usually gets supressed? Did you test on wifi?
-- Dumisa: Have not yet tested with loss-based, will likely increase
base RTT, might cause instability. ensured that we can quickly
measure the RTT.
-- Roland: make sure to test with (artifical) jitter because that
could be challenging for the algorithm.
MOONCAKE: trading more storage for less computation — a KVCache-centric architecture for serving LLM chatbot, paper, Mingxing Zhang, remote, 30 minutes
- Simone: Red Hat uses MOONCAKE as a test case for how LLM systems
behave as an application
- Jana Iyengar: FQ needs per-flow state. Does Santa also have per-flow
state?
-- Ayush: yes, needs per-flow state.
-- Jana: comparison to L4S?
-- Ayush:
- Roland Bless: having a mice queue is a problem, could get problem
with reordering. How do you classify mice flows?
-- Ayush: first 20 packets go into mice queue.
-- Roland: have you looked at slowdown due to reordering?
-- Ayush: have not looked at slowdown, but looked at reordering.
Reordering basically not happening. Order of shuffling has an
impact.
- Stuart Cheshire: +1 to point that too many people prioritize having
more throughput. Latency/throughput not mutually exclusive, good
that the paper also says that. L4S, Prague, etc. should be
considered in more research papers. Comment about "flows lying about
their L4s compliance": L4S not about "better treatment", it's about
getting better informatuion about queue/network state
- Chris Box: comparison to FQ. What is the impact in terms of resource
use. Probably less memory, what about CPU?
-- Ayush: Have built this in P4 switch, so CPU cycles not really
applicable. Regarding memory, not sure that it actually needs less
memory.
- Danesh Zeynali: have you tried this with multiple bottlenecks?
touching the flow on first bottleneck could affect behavior on the
next?
-- Ayush: no, not yet, but interesting dimension.
-- Danesh: what about flows with different RTTs, how do you setting
the fair queuing for these?
-- Ayush: Santa does not care about the RTT / focus on RTT fairness
specifically
- Gorry Fairhurst: look at more queues, not just 1/2, maybe a handful.
complexity to implement this compared to L4S?
-- Ayush: not too complicated to implement but did not check in
terms of L4S
-- Gorry: have you considered impact on "harm"?
-- Ayush: not really focused on harm, yet
- Mohit Tahiliani: have you tested with 90 flows max?
-- Ayush: only tested with 90 flows
-- Mohit: this compares with FairQueuing. There is also FlowQueuing.
FlowQueuing has specific properties, e.g., for sparse flows. Maybe
worth taking that into account
-- Ayush: came across lots of concepts. here, we have a single Mice
queue
-- Mohit: flowqueuing has multiple queues for mice flows with
different priorities
- Michael Welzl: flowqueueing; sparse flows might essentially live
"without a queue". In Santa, these sparse flows might be pushed
together with other flows into one queue.
-- Ayush: have special handling for short/sparse flows but such
parameters could affect the described behavior.
- Jana Iyengar: when mentioned FQ, meant flowqueuing. have you looked
at how applications (with multiple flows) can game the system?
-- Ayush: not yet.