Serverless Performance Modeling with CPU Time Accounting and the Serverless Application Analytics Framework

Abstract

Recently serverless computing platforms have emerged to provide appealing options to developers for deploying cloud native applications. One of the most popular serverless computing delivery paradigms known as Function-as-a-Service (FaaS) offers many desirable features including high availability, fault tolerance, and automatic application scaling. Understanding performance of FaaS workloads in the cloud involves new challenges as FaaS workloads are billed to the nearest millisecond, infrastructure is temporary, and observability of hardware configuration, load balancing, and function tenancy is notoriously obscure. To better understand FaaS workload performance to enable accurate performance predictions, we propose a novel performance modeling approach that leverages Linux CPU Time Accounting (CPU-TA). We introduce the Serverless Application Analytics Framework (SAAF), our tool used to collect information about the infrastructure running FaaS platforms, and use this to train models to predict FaaS workload runtime. Using 10 different serverless applications, we compare our CPU-TA modeling approach to baseline performance modeling approaches that directly predict runtime. Enabling accurate performance predictions on FaaS platforms allows developers and researches to make informed deployment decisions resulting in faster performance and reduced hosting cost. We found our runtime prediction technique was able to make accurate performance predictions across 448 different FaaS configuration scenarios, achieving average prediction accuracy of over 95%.