Session Date/Time: 26 Jul 2024 20:00

iccrg

Summary

The ICCRG meeting in Vancouver focused on congestion control, pacing, and related topics. Presentations covered hackathon results on congestion control testing, slow start threshold after overshoot, pacing in transport protocols, L4S and Prague congestion control, and congestion control response when application limited. Several interesting discussions ensued regarding improving TCP and implementation considerations.

Key Discussion Points

Congestion Control Hackathon: Interoperability testing of "Careful Resume" drafts showed subtle differences between implementations. Future hackathon events are planned with a call for participation.
SSWASH (Slow Start Window After Slow Start High) and Overshoot: Model shows current TCP Reno threshold (0.5) can be too high, especially with pacing, leading to double losses. Suggestions were made to save the congestion window at the time of packet transmission or use BBRv3 techniques to estimate BDP+Q.
Pacing in Transport Protocols: Informational draft aims to document pacing implementations and consequences. Linux TCP, MVFST, and BBR implementations were discussed. Importance of TSO (TCP Segmentation Offload) for fairness was highlighted. The pacer in Google's WebRTC library was mentioned as a possible addition to the draft.
L4S and Prague Congestion Control: Updates on Prague congestion control were presented, including UDP Prague, a standalone congestion control object. Rate compliance, packet pacing, and trade-offs between latency and throughput were discussed.
Congestion Control Response When Application Limited: The problem of setting ssthresh based on flight size after stalls can cause poor performance with streaming applications. It was argued that the current RFC text restricts growth of cwnd even when it is safe to do so. Quick congestion control was mentioned to have addressed this already.

Decisions and Action Items

Pacing in Transport Protocols Draft: Michael Jones will seek contributions from other people to document more implementations.
Congestion Control Response When Application Limited: Further investigation and discussion are needed to understand existing implementations and tradeoffs, possibly leading to a future work item. Matt Mathis will follow up.

Next Steps

Participants interested in the congestion control hackathon in Dublin should contact Reese Enghardt.
Those interested in contributing to the Pacing in Transport Protocols draft should contact Michael Jones.
Continued discussion and investigation of congestion control responses in application-limited scenarios with Matt Mathis.