Credence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
CREDENCE augments commodity drop-tail buffers with ML predictions to emulate push-out behavior, using per-queue thresholds and an “arithmetic push-out” trick (add/subtract sizes) plus a safeguard against false positives. With perfect predictions it matches near-optimal LQD and gracefully degrades to Complete Sharing under heavy prediction error. Packet-level ns-3 eval on a 256-server topology with a random forest predictor shows about 1.5x throughput and up to 95% flow completion time reduction versus prior drop-tail schemes.
AI-written summary · about
chapters (12)
- 0:00Credence introduction
- 1:00Game analogy for buffer sharing
- 2:49Buffer sharing problem
- 3:55Online perspective and competitive ratio
- 5:20Drop tail vs push out
- 6:23Limitations of drop tail
- 7:48Predictions to improve drop tail
- 9:15Goals: robustness and smoothness
- 9:59Credence algorithm design
- 12:12Evaluation and results
- 14:23Future research directions
- 15:25Summary and limitations
related talks
16mPRED: Performance-oriented Random Early Detection for Consistently Stable Performance...
25mApproximating Fair Queueing on Reconfigurable Switches
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
11mScaling Distributed Machine Learning with In-Network Aggregation
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
16mBuilding A CSFQ-Inspired Transport for Switched CXL Memory Pooling
16mFLB: Fine-grained Load Balancing for Lossless Datacenter Networks
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
