Serverless Workload Characteristics

Abstract

Serverless computing and Function-as-a-Service (FaaS) have seen a steady rise, and with more usage, the higher the demand for quick and reliable services becomes. FaaS delivers stress-free scalability and is only billed per execution time for each deployed application for the customer, while the providers retain control over the infrastructure. The objective of this paper focuses on analysing workloads in serverless computing environments with the help of machine learning, to analyse if existing machine learning techniques can be implemented for time-series, or if problems can be overcome. Using machine learning and validation methods, the workloads were analysed, clustered, and characterized in order to identify patterns and offer a way for future improvements to be applied to already existing techniques used by service providers.

The approach taken in this paper is to implement a baseline clustering technique, followed by multiple others to ensure valid results. The clusters are sampled and visualized for characterization, and then grouped based on similar patterns. Multiple common workload patterns are identified across the clusters produced by the defined techniques, which included but are not limited to: Constant, Mean adjacent, Pulse, Sinusoidal, and Spike patterns. Concluding these findings is a discussion about the validity of these common patterns, and suggestions for improvements on these techniques and future studies.