Markdown Version | Session Recording

Session Date/Time: 08 Nov 2024 13:00

iccrg

Summary

The Internet Congestion Control Research Group (ICCRG) meeting featured presentations and discussions on various aspects of congestion control, including pacing in transport protocols, evaluation of BBRv3, speeding up slow start, and an alternative framework for bandwidth allocation. A panel discussion focused on the unique considerations for congestion control in 5G and Wi-Fi networks.

Key Discussion Points

• Pacing and Transport Protocols (michael rosa draft):
  • The draft aims to provide guidance on pacing implementation, showcasing examples from Linux TCP, QUIC BBR, Apple, and FreeBSD.
  • Debugging pacing can be challenging, requiring careful configuration of offload features and accounting for higher-level protocols.
  • Discussion arose regarding a past paper suggesting application-level pacing might be undone by kernel and NIC interactions.
• Evaluation of BBRv3:
  • Systematic evaluation of BBRv3 revealed long convergence times with deep buffers and bias towards longer RTT flows.
  • Coexistence issues with loss-based congestion control algorithms (like Cubic) persist in BBRv3.
  • Enabling ECN did not improve fairness in the tested scenarios.
  • The applicability of BBRv3 for widespread deployment remains questionable due to fairness concerns.
• Speeding up Slow Start (SOS):
  • SOS accelerates congestion window growth in the early phase of TCP connections, aiming to improve bandwidth utilization.
  • The algorithm predicts continuation of exponential growth using a modified version of High Start criteria.
  • A combination of I'm clocking and packet pacing is used to control transmission of additional data packets, minimizing burstiness and packet loss.
  • Experiments suggest that SOS can consistently outperform both BBR and Cubic without compromising fairness.
• Congestion Control Algorithm Independence (CCAI):
  • CCAI aims to decouple throughput from interactions between different congestion control algorithms.
  • The presented framework, Recursive Congestion Shares (RCS), allocates bandwidth based on economic properties of the internet (pay for access).
  • While not universally applicable, RCS provides CCAI in most topologies, including those resembling the internet.
• Panel Discussion: 5G and Wi-Fi Congestion Control:
  • 5G: Resource allocation and channel quality significantly impact congestion control, requiring algorithms to adapt to rapidly changing conditions. Prioritize L4S.
  • Wi-Fi: Features such as collision avoidance (EDCA), aggregation, and half-duplex nature pose unique challenges for congestion control.
  • Packet aggregation at the Wi-Fi layer can improve throughput but introduces a tradeoff with latency and jitter.
  • Discussion explored potential interfaces between the IETF and 802.11 communities for cross-layer collaboration.
  • The general sense was the difficulty in providing congestion control in a more complex and variable medium that requires cross-layer co-engineering.
  • A call for greater Wi-Fi expertise within the IETF was made.

Decisions and Action Items

• Authors of the pacing draft (michael rosa) will incorporate Gory's experience on debugging pacing and a paragraph on the difficulties around it and turn off.
• Amelia to take comments on the start time for staggered experiments in the BBRV3 testing in account and double-check.

Next Steps

• Continue discussions on the BBR draft GitHub issue to refine BBRv3.
• Explore further collaboration between the IETF and IEEE 802.11 communities to address Wi-Fi congestion control challenges.
• Investigate the potential for a joint effort in Pilk.