This paper presents Ursa, a resource management system for cloud-native microservices. Ursa uses an analytical model to break down the end-to-end Service Level Agreement (SLA) into per-service SLA and maps these to individual resource allocations per microservice tier. It addresses the challenges faced by Machine Learning (ML)-driven approaches, such as lengthy data collection processes and limited scalability. In comparison to ML-driven systems, Ursa shortens the data collection process by over 128 times and its control plane is 43% faster. It also reduces the SLA violation rate and CPU allocation during online deployment.
Publication date: 8 Jan 2024
Project Page: https://www.hpca-conf.org/2024/
Paper: https://arxiv.org/pdf/2401.02920
