Flow Scheduling with Imprecise Knowledge
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
QCLIMB uses random forest regression forests to produce per-flow lower and upper bounds (over 99.9% of flows are tightly bounded) and runs a low-bound-based scheduling algorithm that promotes flows based on remaining gap to the lower bound, plus a priority-based loss detection and reordering path to avoid spurious TCP retransmits. It implements inference in the host kernel (≈3 μs per inference), needs only built‑in switch priority queuing, lowers small‑flow average FCT versus PIAS by 88%, and sits about 9% behind pFabric while requiring no switch changes.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 2:22Imprecise flow information
- 6:30Q-ram design overview
- 9:57Loss detection and system
- 11:49Evaluation and experiments
- 15:38Conclusion
related talks
17mIs advance knowledge of flow sizes a plausible assumption?
16mBackpressure Flow Control
25mApproximating Fair Queueing on Reconfigurable Switches
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
28mLoom: Flexible and Efficient NIC Packet Scheduling
16mFLB: Fine-grained Load Balancing for Lossless Datacenter Networks
21mYour Coflow has Many Flows: Sampling them for Fun and Speed
17mPyrrha: Congestion-Root-Based Flow Control to Eliminate Head-of-Line Blocking in...