Abstract: 

Developing cost-efficient serverless DAGs (Directed Acyclic Graphs) is a challenging procedure involving the composition of individual processing steps into serverless functions, determining their file requirements, and optimizing the allocated memory, which in turn affects the networking performance to cloud storage. To address these challenges, we introduce AutoDAG, a file-aware FaaSification and optimization platform aimed at automating the lifecycle of serverless DAGs. AutoDAG automatically composes a fine-grained DAG implementation from a given code, profiles it, and transforms it by fusing subsequent functions, estimating their costs and generating the cost-optimal implementation. AutoDAG estimates the cost of transformed functions with very high accuracy without measuring them. AutoDAG dominates state-of-the-art cost optimizers, both in terms of cost and performance.