Abstract
Serverless computing, specifically the Function-as-a-Service (FaaS) model, provides a scalable foundation for deploying modern applications. However, selecting the optimal memory configuration for AWS Lambda functions remains a challenging task due to the coupling of allocated memory size and CPU share, which creates a non-trivial trade-off between execution duration and monetary cost; classic tuning tools, such as AWS Lambda Power Tuning, ignore the multi-criteria nature of the problem and require an exhaustive search of the configuration space. The aim of the study is to develop an approach for the automated search of Pareto-optimal configurations of FaaS functions without an exhaustive search of the configuration space. The paper proposes an approach to multiobjective optimization based on the non-dominated sorting genetic algorithm NSGA-II with adaptive mutation, memoization of previously measured configurations, and elitist selection, which automatically determines the set of Pareto-optimal configurations. To overcome the performance variation inherent in cloud environments, a robust evaluation strategy has been implemented involving JIT compiler warmup, dropping the first invocation, and median filtering of ten consecutive measurements. Experimentally, on three workloads classes – computational, mixed, and memory-bandwidth-limited – it is demonstrated that the proposed evolutionary approach efficiently converges to a diverse Pareto front, exploring only 23–30% of the search space out of 182 configurations. A "knee point" is consistently identified near the 1769 MB threshold, which corresponds to the allocation of one virtual CPU; cost savings relative to the time-optimal configuration amount to 41–43% for computational workloads, 15–26% for mixed, and 22–39% for memory-bandwidth-limited ones. The scientific novelty lies in the comprehensive combination of NSGA-II with a robust evaluation strategy tailored for cloud environments with unstable performance. The practical significance of the approach is the ability to make informed decisions between high-performance and cost-effective FaaS configurations without exhaustive manual testing, thereby reducing operational expenses for industrial deployments.