A probabilistic modeling and evolutionary optimization approach for serverless workflow configuration

Serverless computing has nowadays become a mainstream paradigm to develop cloud-native applications owing to its high scalability, ease of usage and cost-effectiveness. Nevertheless, because of its poor infrastructure transparency, two main challenges emerge when users migrate their applications to a serverless platform: the lack of an effective analytical model for performance and billing, and the trade-off problem between them. In this paper, we formally define a serverless workflow and introduce the concept of execution instances. Based on them, a probabilistic performance and cost evaluation model is built to obtain their expected values for an input serverless workflow. Then, we design a tailored evolutionary optimization algorithm called EASW to tackle budget-constrained performance optimization and performance-constrained cost optimization problems. Extensive experiments were carried out to test the proposed model and optimization algorithm on AWS Lambda. Results reveal that our model can achieve an accuracy over 98% and EASW can yield a better memory configuration solution than existing methods for constrained optimization.