Efficient Resource Management for Cloud-Native Systems: A Model- Based Optimization Approach

Efficient resource management is a critical challenge in cloud- native systems, demanding a delicate balance between per- formance and operational cost. For microservices, this trans- lates into the complex task of devising an efficient autoscaler that can adapt to dynamic workloads. In second-generation Function-as-a-Service (FaaS) systems, it becomes a difficult trade-off between cost-efficiency and performance degrada- tion, complicated by the manual tuning of the concurrency limit. This thesis develops a unified, model-based optimiza- tion methodology to address these challenges. The approach leverages LQNs, made computationally efficient by a fluid approximation technique. By encoding the resulting ODE sys- tem as constraints within a non-linear optimization problem, we rapidly compute optimal configurations. This methodol- ogy is realized in two distinct frameworks: μOpt, an online autoscaler for microservices that iteratively adjusts CPU and thread-pool resources; and WasteLess, an offline provisioner that performs a single optimization to jointly configure the CPU, memory, and concurrency limit for second-generation FaaS applications. Extensive validation on production-grade cloud platforms proves the methodology’s effectiveness against both industry and state-of-the-art research benchmarks. μOpt improves application performance by up to 9% while reduc- ing resource consumption by up to 30%. For FaaS systems, WasteLess reduces operational costs by an average of 38% against non-concurrent deployments, without performance loss. When compared to the other benchmarks, it lowers la- tency by up to 72%, avoiding the severe performance penalties inherent in their methods.