Lancet: A self-correcting Latency Measuring Tool
Talkat USENIX ATC '19 · on Friday, 2019-07-12· at 10:15· USENIX ATC Track II: Grand Ballroom VII–IX·
summary
Lancet is a self-correcting latency tool that uses NIC hardware timestamping or a busy-polling latency agent, drives experiments with online statistical tests (Anderson‑Darling for inter‑arrival fit, Pearson/correlation checks, Augmented Dickey‑Fuller for stationarity) and automatically adjusts sampling, warmup and duration. Claims: it reports the 99th percentile within ≤10 μs at 95% confidence for the studied workloads, detects inter‑arrival violations (e.g., 15 connections → limit ≈18k rps, 30 → ≈42k rps), and shows ~20 μs lower tails vs Mutilate with NIC.
AI-written summary · about
chapters (13)
- 0:00Start
- 0:53A typical latency experiment
- 2:08Measuring tail-latency is challenging
- 2:58Open versus closed queueing
- 3:56Inter arrival distribution impact
- 6:04Lancet workflow
- 8:06Experiment Decomposition
- 9:53Lancet Design
- 11:04Outline
- 13:36Identifying inter-arrival distribution violations
- 15:30Sample count for target CI
- 16:16Conclusion
- 17:57Lancet in action
related talks
17mHow to diagnose nanosecond network latencies in rich end-host stacks
43mHow NOT to Measure Latency
01h 01mThe Performance Investigator's Field Guide
13mMeasuring Real-Life Latency of the Internet: A Netflix Story
44mWhy Do Clients Nitpick?
12mHistory-based Latency Prober Tuning
37mLies, Damn Lies, and Metrics
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...