PerfDB: A Data Management System for Fine-Grained Performance Anomaly Detection
Joshua Kimball, Rodrigo Alves Lima, Yasuhiko Kanemasa, Calton Pu · 2020
In this work, we present our performance data management system, PerfDB, that we use to study fine-grained performance anomalies like Millibottlenecks. We use it to present the first experimental evidence of a phenomenon we call, “Localized Latency Requests.” These are performance bugs that are part of the long-tail of system latency. We also provide a population study of Very Long Response Time (VLRT) requests, a separate performance anomaly belonging to the latency long tail, being inducing by millibottlenecks through queueing effects.