Utilization-focused Evaluation on Serverless Architectures

Heron Yang
Kung-Hsien Yu

ABSTRACT

Serverless computing is a pay-as-you-go code execution platform in
which the cloud provider fully manages the execution environment to
serve requests. According to one study, adopting serverless
architectures decreases the application operating cost to 77.08%,
which leads its popularity in ap plication developer communities. A
sufficient understanding of the system utilization of a serverless
architecture will benefit both service providers in designing
architectures and application developers in making adoption decisions.
In this paper, we first describe design goals of a serverless
architecture and then evaluate three serverless architectures:
OpenLambda, IronFunctions, and Clofly. Both OpenLambda and
IronFunctions use Docker containers with different state operations on
containers for providing application isolation, while Clofly handles
requests purely by using subprocess calls. We observe that the current
isolation solution costs five times more on server resource usages
comparing to the one with no isolation guarantee. Also, we discuss the
importance of limiting resource allocations for user functions,
scalability-concerned cache policies, and efficiency improvements by
sharing runtime resources.