Cost-optimal Operation of Latency Constrained Serverless Applications: From Theory to Practice.

Serverless computing and the function as a service model are new paradigms enabling the fine granular, bottomup construction of cloud-native applications. It can significantly reduce operating costs while shifting the management tasks from developers and application providers towards the cloud operators. But these benefits are provided at the cost of less control over the underlying infrastructure and the application performance, including the end-to-end latency. However, grouping of functions into deployable serverless software artifacts remains still under our control, which has a considerable impact on performance and operation costs. In this paper, we propose fast and efficient algorithms that can partition an application’s functions into separate deployment artifacts in a cost-optimal way while meeting user-defined average end-to-end latency bounds. Moreover, our approach supports the dynamic redesign and reconfiguration of the current deployment setup in response to changes in monitored metrics. Our main contribution is threefold. First, we establish the relevant theoretical models capturing the behavior of the serverless ecosystem and we define the main problem. In addition, the concept of the integrated application management is introduced. Second, we propose novel algorithms providing optimal solutions for different variants of the core problem and the complexity of the methods are analyzed. Third, we demonstrate the applicability and the benefits of our solution by evaluating different deployment scenarios of a realistic use case in Amazon’s public cloud environment.