Keyun Cheng

SketchLearn: Relieving User Burdens in Approximate Measurement with Automated Statistical Inference

Download

SIGCOMM, 2018

Summary

Approximate network measurements trade accuracy to saves resources, but requires intensive manual effort to learn the appropriate tradeoff. Sketchlearn, a sketch-based network measurement framework, learns statistical properties from the resources to eliminates the resource traffic conflicts.

Details

Problems behind: Due to conpetition of network traffics with limited resources (resource conflicts), measurement errors occurs. Sufficient resources must be guaranteed in approximate measurements. A tight binding of resource configurations (resource params) and accuracy parameters exists in approximate measurements.

Current approach: approximate measurements. Limitations: Hard to quantify stats (expected errors, threshholds), hard to find theoretical bounds, hard to examine correctness. Not understand: Hard to define flowkeys

Design requirements: fast (real-time, per-packet processing), resource saving (memory), generalization

Proposed approach: Sketch-based measurement framework. It draws the resource conflicts by building a multi-level sketches. The sketch tracks the frequencies of flow records at bit-level, so the multi-sketch forms a multi-level Gaussian.

Model learning (needs to look in detail): Theory about model inference. Large flow extraction replies on this.

Implementation: Software/hardware data plane and control plane. Flows are updated with each incoming packet.

Strength

Weakness