Explore efficient sketch-based network monitoring on programmable switches in this NSDI '22 conference talk. Delve into the challenges of implementing sketching algorithms on hardware and discover SketchLib, a library designed to optimize resource usage. Learn about systematic bottleneck analysis, practical optimization techniques, and the benefits of using SketchLib to reduce hardware resource footprint by up to 96% without compromising accuracy. Gain insights into consolidating hash calls, removing unnecessary SALU allocations, and implementing end-to-end code examples. Understand the applicability of SketchLib and its impact on preserving sketch accuracy while significantly reducing resource bottlenecks in programmable switch environments.
SketchLib - Enabling Efficient Sketch-based Monitoring on Programmable Switches
Overview
Syllabus
Intro
Let's try to run sketches on programmable switches
Sketch is inefficient or infeasible on programmable switch
Our work: Sketchlib bridges the gap!
Outline
Two background questions for bottleneck analysis
Sketches have four common steps
Sketch imposes burden on hardware resources
Hash calls and SALUs are resource bottlenecks
Overview of optimizations and SketchLib API
01. Consolidate short hash calls
05. Remove unnecessary SALU allocation
End-to-end code example
Applicability of SketchLib
Sketchlib reduces resource bottlenecks
Sketch Lib preserves accuracy
Conclusion
Taught by
USENIX
