OSCAR: O(1)-Step Convergence and Readily-deployable Congestion Control
summary
Claims O(1)-step convergence using end-to-end signals rather than INT, implements MIMD updates by estimating delay gradient via batch slope fitting with a constant time/space formula suitable for NICs, and coordinates separate window and rate control loops into a single multiplicative factor. Testbed shows convergence in one to two update steps; CPU overhead is slightly above DCQCN but below other competitors. Large-scale sims report 12–48% average FCT and 40–74% tail FCT improvement versus precise-INT-based CCs.
AI-written summary · about
chapters (13)
- 0:00Introduction to OSCAR congestion control
- 1:22O(1)-step convergence need
- 3:36O(1)-step convergence defined
- 6:24Limitations of deployable solutions
- 7:29Multiplicative increase/decrease
- 9:16Delay and gradient signals
- 11:48OSCAR design challenges
- 13:11Synchronized state selection
- 14:35Window and rate coordination
- 16:03Fairness and implementation
- 16:24Implementation and overhead
- 17:26Swift vs H-TCP vs OSCAR
- 18:30Conclusion
related talks
19mSwCC: Software-Programmable and Per-Packet Congestion Control in RDMA Engine
24mCopa: Practical Delay-Based Congestion Control for the Internet
17mTowards provably performant congestion control
17mBridging the Gap between QoE and QoS in Congestion Control: A Large-scale Mobile...
18mRevisiting Congestion Control for Lossless Ethernet
16mPRED: Performance-oriented Random Early Detection for Consistently Stable Performance...
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
17mRe architecting Congestion Management in Lossless Ethernet