Using application knowledge to reduce cold starts in FaaS services

Abstract

In Function-as-a-Service platforms (FaaS), which have become very popular lately, code is deployed in the unit of single functions and the cloud provider handles resource management. There, a key problem is the so-called cold start problem: when a request comes in and no idle container can be found for the execution of the target function, then a new container needs to provisioned. In that case, the request incurs an extra latency - the cold start latency. Recent work has largely focused on reducing the duration of cold starts. In this paper, we present three approaches, complementary to related work, that reduce the number of cold starts while treating the FaaS service as a black box. In the approaches, implemented as part of a lightweight choreography middleware, we use knowledge on the composition of functions to trigger cold starts and, thus, the provisioning of new containers before the application process invokes the respective function. In experiments on AWS Lambda and OpenWhisk, we could show that our approaches remove an average of about 40% (in some cases up to 80%) of all cold starts while causing only a small cost overhead.