Data-intensive task scheduling in geo-distributed cloud computing: a deep reinforcement learning approach

Geo-distributed clouds have become an increasingly common architecture in modern cloud computing. By placing data centers in multiple geographical regions, cloud service providers shorten the distance between users and cloud resources, enabling low-latency service delivery worldwide. However, this architecture raises new challenges for scheduling data-intensive tasks, including limited inter-regional bandwidth and high traffic costs charged by Internet service providers. These factors can significantly impact both performance and cost efficiency. Existing scheduling approaches for geo-distributed clouds often rely on mathematical and meta-heuristic methods. They typically focus on batch tasks, unable to make real-time decisions and exhibit limited adaptability to changing workload patterns. To address these constraints, this paper introduces a deep reinforcement learning-based framework for scheduling data-intensive tasks across geographically distributed data centers. The framework performs real-time scheduling while considering bandwidth constraints and traffic costs, enabling multi-objective optimization of task response time and total operational cost. Extensive experiments under various workloads show that the proposed method achieves competitive and often better performance than the evaluated baselines.