TL;DR
Sensitivity and ablation studies confirm stable learning and controllable latency-cost trade-offs, demonstrating that lightweight RL can effectively deliver cost-efficient, adaptive autoscaling in hybrid cloud environments.
Cloud software systems operate under highly dynamic and unpredictable workloads, requiring efficient and adaptive resource management strategies to maintain performance, reliability, and cost efficiency. Traditional rule-based and heuristic resource allocation approaches often fail to respond optimally to rapid workload fluctuations and complex system interactions. This paper proposes a reinforcement learning-based adaptive resource management framework that enables cloud systems to autonomously learn optimal resource allocation policies through continuous interaction with the environment. By modeling cloud resource management as a sequential decision-making problem, the framework leverages reinforcement learning algorithms such as Q-learning, Deep Q-Networks (DQN), and policy-gradient methods to dynamically adjust computing resources including CPU, memory, and virtual machine instances. The proposed approach aims to optimize multiple objectives such as performance, cost, and service-level agreement (SLA) compliance. Experimental evaluation using simulated and real-world cloud workloads demonstrates that reinforcement learning significantly outperforms static and reactive baseline strategies in terms of resource utilization efficiency and response time stability. The results highlight the potential of reinforcement learning to enable intelligent, self-adaptive cloud resource management systems.
Cloud computing environments face increasingly complex challenges in task scheduling due to dynamic workloads, heterogeneous resources, and multi-objective optimization requirements. This paper proposes an innovative Deep Reinforcement Learning (DRL)-based Intelligent Task Scheduling Framework (DRITS) designed to optimize task allocation and resource utilization in cloud distributed systems. The proposed framework leverages advanced Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO) algorithms to enable dynamic, adaptive scheduling that continuously learns optimal policies through interaction with the cloud environment. Our comprehensive evaluation demonstrates that DRITS achieves significant performance improvements, including 32.4% reduction in makespan, 48.7% lower energy consumption, and 22.6% improvement in resource utilization compared to traditional heuristic algorithms [1]. Extensive simulations using real-world Google Cluster workloads and diverse benchmark datasets validate the robustness and scalability of the proposed approach across varying workload conditions. The framework demonstrates strong adaptability to dynamic environments, fault tolerance capabilities, and superior performance in multi-objective optimization scenarios. These results establish DRL-based intelligent scheduling as a promising solution for next-generation cloud computing infrastructure management.
Cloud computing is a key enabler of modern computing services, offering scalability and flexibility. However, efficient management of cloud resources remains challenging due to limited capacity and the increasing number of tasks requiring timely execution. An effective task scheduling strategy is therefore essential to improve resource allocation and utilization, reduce operational costs and energy consumption, and support high availability-especially for long-term jobs. In this paper, we propose a new scheduling approach that combines a Social-Based Algorithm (SBA) with Deep Reinforcement Learning (DRL), referred to as SBA-DRL. This method allocates tasks to resources by learning from workload patterns and adapting to workload characteristics in a batch scheduling context. We evaluate SBA-DRL using both a synthetic dataset and the real-world Google Cloud Jobs (GoCJ) under workloads ranging from 200 to 1,000 tasks. On the synthetic dataset, our method reduces cost by 20.21% and energy consumption by 25.31%, while improving resource utilization by 9.36%. On the GoCJ dataset, it achieves up to 28.94% lower cost, 8.16% less energy use, and a 14.04% increase in resource utilization. In both cases, SBA-DRL also demonstrates better performance in resource allocation and high-availability management compared to existing heuristics, meta-heuristics, hybrid, and machine learning-based schedulers. These results indicate that the proposed SBA-DRL approach effectively addresses key challenges in cloud task scheduling, offering a practical solution to enhance the efficiency and sustainability of cloud systems.
In today’s world, with the emergence of IoT devices and critical latency applications, there has been an increased demand for intelligent resource management in diverse computing environments [5][10]. Collaborative computing is the integration of cloud and edge computing, while the issue of how to schedule each task for execution is still an open problem. Traditional heuristics like Round Robin and greedy latency reduction are incapable of adapting to the stochastic and non-stationary characteristics of practical workloads [4][14]. In this paper, an intelligent task scheduling mechanism based on Double Deep Q-Network (DDQN) reinforcement learning [2] has been proposed. The agent observes a four-dimensional state encoding task characteristics and selects binary offloading decisions, guided by a shaped reward signal encoding multiple performance objectives. Experimental evaluation on a heterogeneous synthetic benchmark demonstrates that the proposed DDQN scheduler reduces SLA violations by approximately 85% relative to Round Robin and 72% relative to the greedy baseline, while achieving superior energy efficiency. These results confirm that deep reinforcement learning [1][17][18] provides a principled foundation for adaptive resource management in next-generation edge-cloud systems.
Large Language Model (LLM) inference services are increasingly deployed across multi-cloud infrastructures to support latency-sensitive applications such as conversational AI, intelligent search, and real-time analytics. However, the coexistence of heterogeneous computing resources, geographically distributed data centers, and highly dynamic request patterns poses significant challenges to efficient L7 load balancing and global traffic scheduling. In particular, traditional rule-based or heuristic-driven approaches fail to capture the semantic variability of LLM workloads, including token length diversity and non-linear inference latency, leading to suboptimal resource utilization and degraded tail latency performance. To address these issues, this paper proposes H2-LBM, a hierarchical hybrid deep reinforcement learning framework designed for multi-cloud LLM serving environments. The framework decomposes the scheduling problem into two coordinated levels: a global scheduler based on Proximal Policy Optimization (PPO) for cross-cloud traffic allocation, and a local scheduler based on Double Dueling Deep Q-Network (D3QN) for fine-grained L7 request dispatching. By incorporating semantic-aware state representations and a multi-objective reward function that jointly optimizes latency, throughput, and resource efficiency, H2-LBM enables adaptive and scalable decision-making under dynamic workloads. Experimental results on a multi-cloud Kubernetes testbed show that H2-LBM reduces P99 latency by 21.5%–29.2% compared with strong baselines such as Kubernetes HPA across moderate and burst workloads. Specifically, P99 latency decreases from 360 ms to 255 ms under moderate load and from 650 ms to 510 ms under high load. These results consistently validate the effectiveness of H2-LBM in improving latency stability and system efficiency for large-scale LLM inference services.