Multi-Objective Deep Q-Learning for Joint Cost Mitigation, Thermal Efficiency, and SLA Reliability in Serverless Cloud Infrastructure

• Billy Elly

  Lautech

  Author

Serverless computing has emerged as a dominant cloud service model, offering fine-grained resource allocation and pay-as-you-go pricing. However, the dynamic and ephemeral nature of serverless functions introduces significant challenges in simultaneously optimizing operational costs, thermal efficiency, and Service Level Agreement (SLA) compliance. Existing heuristic-based schedulers often prioritize single objectives, leading to suboptimal trade-offs that either increase carbon footprint or degrade user experience. This study addresses this gap by proposing a Multi-Objective Deep Q-Learning (MO-DQN) framework that jointly optimizes cost mitigation, thermal efficiency, and SLA reliability in serverless cloud infrastructures. The framework models function scheduling as a sequential decision-making problem, with a reward function designed to balance monetary cost, energy consumption, and latency SLO violation rates. Experimental evaluations on a simulated serverless environment using Azure Functions trace data demonstrate that the proposed MO-DQN framework achieves a 34.6% reduction in operational cost, a 22.3% improvement in thermal efficiency, and an 18.7% decrease in SLO violation rates compared to heuristic baselines. The framework's performance approximates optimal solutions within a factor of 1.08 while reducing scheduling time by 99%. The findings provide a replicable framework for cloud providers seeking sustainable and performance-aware resource management, with practical implications for green cloud computing and SLA-driven autoscaling.

• pdf

Copyright (c) 2026 Billy Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.