Publication
CLOUD 2022
Conference paper

MicroLens: A Performance Analysis Framework for Microservices Using Hidden Metrics With BPF

Abstract

Determining the root cause of performance regres-sion for microservices is challenging. The topological cascadingperformance implications among microservices hide the sourceof the problem. Additionally, the lack of knowledge about appli-cation phases can potentially lead to false-positive critical servicedetection. Service resource utilization is an imperfect proxy forapplication performance, potentially leading to false positives.Therefore, in this work, we propose a new performance testingframework that leverages hidden Berkeley Packet Filter (BPF)kernel metrics to locate root causes of performance regression.The framework applies a systematic multi-level approach toanalyze microservice performance without intrusive code instru-mentation. First, the framework constructs an attributed graphwith microservice requests, scores the services to identify thecritical paths, and ranks the low-level metrics to highlight the rootcause of performance regression. Through judiciously designedexperiments, we evaluated the metric collection overhead, show-ing less than 18% more latency when the application is runningacross hosts and 9% within the same host. In addition, dependingon the application, no overhead is experienced, while the state-of-the-art approach presented up to 1060% more latency. Themicroservice benchmark evaluation shows that MicroLens cansuccessfully identify the set of root causes and that the causesvary when the application is running in different infrastructures.