PRED: Performance-oriented Random Early Detection for Consistently Stable Performance...
summary
PRED implements on- switch flow counting and two loosely coupled subsystems, Flow Concurrent Stabilizer (FCS) and Queue Length Adjuster (QLA), using MIMD to scale RED slope with active-flow concurrency and A/B-style trials plus a utility function to nudge RED safely for traffic mix. Prototype runs on Tofino1/2 hardware with DCTCP hosts; testbed and large-scale sims show 66% lower switch queue length and up to 80% lower median FCT versus static RED, and 34% tail FCT reduction versus a learning baseline.
AI-written summary · about
chapters (9)
- 0:00Context and motivation
- 0:52Congestion control methods
- 3:11RED parameter optimization
- 5:14Traffic-aware insights
- 8:48Design of PRED
- 10:25PRED architecture
- 12:37Implementation details
- 13:19Evaluation and results
- 14:25Conclusions
related talks
16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
25mApproximating Fair Queueing on Reconfigurable Switches
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
16mFLB: Fine-grained Load Balancing for Lossless Datacenter Networks
16mRuntime Programmable Switches
16mBackpressure Flow Control
19mTackling DC Congestion and Bursts
19mOSCAR: O(1)-Step Convergence and Readily-deployable Congestion Control
