OctoSketch: Enabling Real-Time, Continuous Network Monitoring over Multiple Cores
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
OctoSketch scales sketches across many cores by sending only counters that change enough to matter, using continuous change-based aggregation. It strips heavy per-core structures and removes per-core flow storage, dynamically adjusts a universal threshold to match packet rates, and runs on CPU, DPDK and eBPF XDP. Implementations on nine representative sketches claim 15.6× lower errors and up to 4.5× higher throughput versus full-sketch merging; source code is not available.
AI-written summary · about
chapters (9)
- 0:00OctoSketch introduction
- 1:42Sketching network statistics
- 3:49Scaling sketching to multicore
- 5:39Entire sketch merge limits
- 7:03OctoSketch change-based mechanism
- 8:48Dynamic resource allocation
- 10:34Removing redundant structures
- 11:16Evaluation results
- 13:24Conclusion and future work
related talks
15mSketchPipe: Toward Accurate Sketch-based Network Measurement on Multi-Pipeline...
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
20mCan OVS-DPDK be further optimised?
12mToward Nearly-Zero-Error Sketching via Compressive Sensing
36mThe black art of network performances