Skip to content

Congestion Control Deep Dive · Part 9 of 9

Chapter Summary

Jump to section

1. Why Congestion Control Is the Most Consequential Knob in a GPU Cluster2. DCQCN: The Algorithm Powering Most Deployed AI Fabrics3. Swift: RTT-Based CC Without ECN4. HPCC: INT-Powered Precise Congestion Control5. TIMELY: RTT-Gradient CC on Commodity NICs6. UEC CC Revisited: The 1-RTT Architecture7. Comparison Table and Algorithm Selection Guide8. NIC Firmware Implementation and Parameter Tuning Reference9. Chapter Summary

Congestion control in AI fabrics is an algorithm-level competition measured in RTTs. The key insight across all five algorithms:

  • DCQCN: 2-RTT feedback via ECN → CNP → rate reduce. Deployed on every ConnectX-7. Requires switch ECN thresholds and CNP DSCP trust to be correctly configured.
  • Swift: 1-RTT feedback via RTT measurement at NIC. Excellent under incast. Requires custom NIC hardware. Not available on ConnectX-7.
  • HPCC: 1-RTT feedback via INT headers from every switch hop. Most precise. Requires INT-capable switch ASIC. High operational complexity.
  • TIMELY: Reacts to RTT gradient, not absolute RTT. Works on commodity NICs. Practical only when baseline RTT >> 20 µs (noise floor constraint).
  • UEC CC: 1-RTT feedback via C-flag in ACK. No separate CNP. Optional NPM telemetry. Available today on AMD Pensando Salina; NVIDIA future roadmap.

For the overwhelming majority of deployed NVIDIA GPU clusters today: get DCQCN right. The difference between a correctly tuned DCQCN deployment and a misconfigured one is 20–40% JCT improvement. Labs 12 and 13 give you hands-on practice diagnosing the most common DCQCN misconfiguration patterns and benchmarking algorithm behaviour directly.

100% complete

Previous

NIC Firmware Implementation and Parameter Tuning Reference

Chapter complete

Back to curriculum