Evaluating Performance of Serverless Virtualization

Abstract

The serverless computing has posed new challenges for cloud vendors that are difficult to solve with existing virtualization technologies. Maintaining security, resource isolation, backwards compatibility and scalability is extremely difficult when the platform should be able to deliver native performance. This paper contains a literature review of recently published results related to the performance of virtualization technologies such as KVM and Docker, and further reports a DESMET benchmarking evaluation against KVM and Docker, as well as Firecracker and gVisor, which are being used by Amazon Web Services and Google Cloud in their cloud services.

The context for this research is coming from education, where students return their programming assignments into a source code repository system that further triggers automated tests and potentially other tasks against the submitted code. The used environment consists of several software components, such as web server, database and job executor, and thus represents a common architecture in web-based applications.

The results of the research show that Docker is still the most performant virtualization technology amongst the selected ones. Additionally, Firecracker and gVisor perform better in some areas than KVM and thus are viable options for single-tenant environments. Lastly, applications that run untrusted code or have otherwise really high security requirements could potentially leverage from using either Firecracker or gVisor.