Investigating Performance and Cost in Function-as-a-Service Platforms

Abstract

The function-as-a-service (FaaS) service model has been gaining adoption at a fast pace. In this service model, cloud applications are structured as self-contained code modules called functions that are instantiated on-demand, and billing is based on the number of function invocations and on function execution time. Developers are attracted to FaaS because it promises to remove two drawbacks of the traditional IaaS and PaaS service models, the need to provision and manage infrastructure, and the need to pay for unused resources. In practice, however, things are a little less rosy: developers still have to choose the amount of memory allocated to functions, and costs are less predictable, especially because they are tied to function performance. This work investigates performance and cost variations within and across FaaS providers. Our results show that performance and cost can be significantly affected by the choice of memory allocation, FaaS provider, and programming language: we observed differences of up to $8.5\times$ in performance and 67$\times$ in cost between providers (with the same language and memory size), and $16.8\times$ in performance and $67.2\times$ in cost between programming languages (with the same provider and memory).