OpenDC Serverless: Design, Evaluation, and Implementation of a FaaS Platform Simulator

Abstract

Function-as-a-Service (FaaS) pertains to a recent advancement in cloud computing known
as serverless computing. A group of technologies provided as ”services” that shift the
responsibility of provisioning resources to the cloud operator and offer a fine-grained cost
model. Despite the growing popularity of FaaS within the research community, evaluating
the performance and cost of different resource management, scheduling, and provisioning
policies remains a difficult endeavor. Conducting experiments in the cloud is costly and
usually yields unpredictable results due to the underlying hardware heterogeneity of cloud
infrastructures. Whereas previous work has focused on providing open-source FaaS im-
plementations, the requirements for conducting a custom practical systems study in FaaS
remain steep both intellectually and financially.
To render practical systems research in FaaS more accessible, we propose in this work
OpenDC Serverless: an expandable trace-based simulator that provides a toolkit for mod-
elling and testing custom FaaS patterns. The simulator exposes custom interfaces for the
implementation of resource allocation, management, and scheduling policies. It further
supports the modification of its core architectural components. Moreover, included in the
toolkit is an experimentation framework that specifies an easy-to-convert-to trace format,
facilitating the exchange of inputs in the community. OpenDC Serverless also provides an
interface for conducting controlled and repeatable experiments. It allows checking each
experiment with fine granularity, through a metric monitoring system that allows tracking
of detailed metrics in the simulation.
We prototype OpenDC Serverless and conduct with it representative experiments. We
showcase the simulation performance of OpenDC serverless in a reproduction experiment
that involves the optimisation of a real FaaS workload through the use of a custom op-
erational policy; The results indicate that the simulator produces valid results. We then
highlight the ability of OpenDC Serverless to explore and characterize real scenarios with
an experiment on simulating cold-starts in AWS Lambda.