SketchPipe: Toward Accurate Sketch-based Network Measurement on Multi-Pipeline...
summary
Places each sketch array exclusively on one pipeline with splitless sketch placement and asynchronously ships cached flow keys via generated state packets so normal packets do not recirculate. Implements P4 data plane plus a C++ placer, runs on 12.8 Tbps Tofino2 hardware, and claims accuracy gains up to 81% and error reductions up to two orders of magnitude. Reported overhead is small: pipeline usage <10%, buffer usage <1%, recirculation bandwidth ~0.2%.
AI-written summary · about
chapters (14)
- 0:00SketchPipe introduction
- 1:28Sketch placement challenges
- 4:14Splitless sketch placement
- 6:45Resource allocation optimization
- 7:29Caching for flow tracking
- 8:55State packets handling
- 9:41Implementation and evaluation
- 10:22Accuracy improvements
- 11:48Multiple monitoring apps
- 12:31Robustness under imbalance
- 12:52Scalability with pipelines
- 13:34Resource efficiency
- 13:57Performance overhead
- 14:19Summary
related talks
14mOctoSketch: Enabling Real-Time, Continuous Network Monitoring over Multiple Cores
16mHeteroSketch: Coordinating Network-wide Monitoring in Heterogeneous and Dynamic Networks
17mSketchovsky: Enabling Ensembles of Sketches on Programmable Switches
15mSketchLib: Enabling Efficient Sketch-based Monitoring on Programmable Switches
22mSoftware Defined Traffic Measurement with OpenSketch
12mToward Nearly-Zero-Error Sketching via Compressive Sensing
15mAutoSketch: Automatic Sketch-Oriented Compiler for Query-driven Network Telemetry
16mState-Compute Replication: Parallelizing High-Speed Stateful Packet Processing