Adaptive Resource Scheduling for Large-Scale Generative AI Models in Cloud Environments

Review Adaptive Resource Scheduling for Large-Scale Generative AI Models in Cloud Environments Elias Dritsas * and Maria Trigka Academic Editors: Muhammad Kazim, Mujeeb Rehman and Stefan Bruce Received: 15 December 2024 Revised: 8 February 2025 Accepted: 11 February 2025 Published: 17 February 2025 Copyright: © 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/ licenses/by/4.0/). Industrial Systems Institute (ISI), Athena Research and Innovation Center, 26504 Patras, Greece; [email protected] * Correspondence: drit[email protected] Abstract: The emergence of large-scale generative AI models, including transformer-based and diffusion architectures, has introduced significant challenges in efficient resource management within cloud environments. Conventional static or heuristic scheduling strategies are often inadequate for handling the dynamic, heterogeneous, and computation-intensive nature of these workloads. This study presents an adaptive resource scheduling framework designed to optimize the allocation of compute, storage, and networking resources for generative AI workloads in multi-cloud infrastructures. The proposed approach employs machine learning–based prediction models and real-time performance analytics to enable dynamic scaling, intelligent load distribution, and proactive resource provisioning. Experimental results using representative large-model benchmarks demonstrate that the framework achieves up to 35% reduction in resource underutilization and 25% improvement in inference throughput compared with traditional scheduling methods. Additionally, the framework supports energy-aware scaling, reducing operational costs and improving sustainability. The findings underscore the effectiveness of adaptive scheduling as a key enabler for high-performance, cost-efficient, and environmentally sustainable deployment of large-scale generative AI models in cloud computing environments. Keywords: Generative Artificial Intelligence (AI) Transformer and Diffusion Models Cloud Resource Management Adaptive Resource Scheduling 1. Introduction 1. Introduction 1.1 Background and Motivation The exponential growth of generative artificial intelligence (AI) has transformed the landscape of modern computing. Models such as large language models (LLMs), transformer-based architectures, and diffusion models have achieved unprecedented capabilities in text generation, image synthesis, and multimodal understanding. However, these advancements come with immense computational, storage, and networking demands, particularly during large-scale training and inference phases. The complexity and size of these models—often involving billions of parameters—require specialized hardware such as GPUs and TPUs, coupled with scalable data pipelines and distributed compute environments. Cloud computing has emerged as the dominant infrastructure for deploying and managing these workloads, offering elasticity, scalability, and on-demand provisioning of highperformance resources. Major cloud providers now support AI-specific services optimized for generative workloads. Yet, the inherent variability of AI model execution, fluctuating user requests, and resource contention across virtualized clusters introduce critical challenges in maintaining optimal performance, cost-efficiency, and sustainability. Big Data Cogn. Comput. 2025, 9, 44 2 of 10 1.2 Problem Statement Despite advances in cloud resource orchestration, efficient scheduling of compute, memory, and network resources remains a major bottleneck in large-scale generative AI operations. Traditional static or rulebased scheduling algorithms are ill-suited to the highly dynamic nature of these workloads, where computational demand and resource availability change unpredictably. These limitations often result in underutilized hardware, increased latency, energy inefficiency, and escalating operational costs. Moreover, as generative AI applications expand across multi-cloud and hybrid environments, existing schedulers struggle to balance load distribution, data locality, and real-time adaptability. 1.3 Research Objectives This research aims to address the above challenges through the development of an adaptive resource scheduling framework tailored for large-scale generative AI workloads in cloud environments. The specific objectives are: 1. To design adaptive scheduling mechanisms that dynamically allocate compute, memory, and networking resources based on realtime workload behavior and performance metrics. 2. To enhance the performance, scalability, and energy efficiency of generative AI training and inference pipelines. 3. To evaluate the proposed framework using representative large-model benchmarks and compare it with traditional cloud schedulers. 1.4 Significance of the Study The proposed study holds substantial relevance for cloud service providers, AI developers, and data center managers. By introducing adaptive, data-driven scheduling techniques, this research contributes to the broader goal of sustainable and efficient AI infrastructure management. It provides actionable insights into optimizing large-scale AI deployments, minimizing energy consumption, and improving the overall quality of service (QoS). Furthermore, the study’s outcomes align with global efforts toward green computing and intelligent cloud orchestration, positioning adaptive scheduling as a foundational technology for the next generation of AI-driven cloud ecosystems. 2. Literature Review 2.1 Resource Scheduling in Cloud Environments Resource scheduling has long been a fundamental concern in cloud computing, directly influencing system performance, cost efficiency, and service reliability. Traditional approaches—such as First Come First Serve (FCFS), Round Robin (RR), and MinMin/Max-Min algorithms—were initially effective for homogeneous workloads but lack the flexibility required by complex AI-driven tasks. Heuristic-based schedulers, including Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO), have been widely adopted to improve resource utilization and minimize execution time (Li et al., 2022). However, these methods typically rely on static parameters or pre-defined rules, limiting their ability to respond to dynamic workload fluctuations. In contemporary multi-tenant cloud environments, where workloads vary in computational intensity and data flow, these traditional schedulers often struggle to maintain load balance and fairness across virtual machines or containers. 2.2 AI Workload Management Techniques The rise of deep neural networks (DNNs) and large language models (LLMs) has prompted significant research into workload management strategies tailored for AI applications. Early frameworks such as TensorFlow Serving and TorchServe focused primarily on inference optimization rather than holistic resource scheduling. Later developments, including Kubernetes-based AI orchestration systems, introduced autoscaling and GPU-aware scheduling mechanisms to better support AI training and inference (Zhao et al., 2023). Despite these advancements, most AI workload schedulers are designed for static or predefined training configurations, making them less effective in handling the stochastic and adaptive nature of generative AI models. Moreover, large-scale generative workloads Big Data Cogn. Comput. 2025, 9, 44 3 of 10 exhibit temporal variability—characterized by unpredictable token generation, variable sequence lengths, and irregular GPU memory utilization—that conventional schedulers fail to accommodate efficiently. 2.3 Adaptive and Dynamic Scheduling Approaches To overcome the limitations of static scheduling, researchers have explored adaptive and feedback-driven mechanisms. Reinforcement Learning (RL) has emerged as a promising paradigm for self-optimizing cloud scheduling, where the scheduler learns from historical performance metrics to make dynamic resource allocation decisions (Kumar et al., 2024). Other approaches leverage feedback control systems, predictive analytics, and machine learning-based forecasting to adapt scheduling policies in real time. Additionally, auto-scaling frameworks—such as AWS Auto Scaling and Kubernetes Horizontal Pod Autoscaler—enable elasticity by provisioning additional resources based on workload metrics. However, these systems typically react to coarsegrained signals (e.g., CPU or memory utilization) rather than model-specific demands, which limits their applicability to generative AI pipelines that require fine-grained adaptation across GPUs, interconnect bandwidth, and memory hierarchies. 2.4 Gaps and Research Opportunities While adaptive scheduling techniques have advanced considerably, significant research gaps remain in the context of large-scale generative AI workloads. Current solutions often lack real-time adaptation mechanisms capable of responding to rapid changes in inference and training dynamics. Furthermore, most frameworks fail to integrate crosslayer optimization that jointly considers computation, storage, and network constraints. In multi-tenant and heterogeneous cloud environments, the coexistence of diverse workloadsranging from conventional DNN training to real-time generative inference— creates resource contention and performance isolation issues that existing schedulers do not adequately address. There is a clear need for intelligent, workload-aware, and energy-efficient scheduling frameworks that can dynamically optimize resource allocation for generative AI models without compromising scalability or cost efficiency. This gap forms the basis for the present study, which proposes an adaptive resource scheduling framework leveraging predictive modeling and real-time feedback to enhance performance and sustainability in large-scale cloud-based generative AI systems. 3. System Architecture and Design 3.1 Cloud Environment Setup The proposed adaptive scheduling framework is deployed within a cloud-native environment composed of virtualized and containerized compute clusters. The infrastructure integrates virtual machines (VMs) and containers orchestrated through Kubernetes or equivalent container management platforms. Each cluster node hosts heterogeneous hardware resources, including GPUs, CPUs, and high-speed NVMe storage, interconnected through high-bandwidth network fabrics such as InfiniBand or 100 Gb Ethernet. The distributed architecture allows the dynamic provisioning of resources across multiple availability zones and data centers, supporting both training and inference workloads for generative AI models. Resource allocation is managed through an underlying cloud management layer, which interfaces with the adaptive scheduling framework via APIs for resource discovery, telemetry, and task orchestration. This setup ensures portability and scalability across multi-cloud or hybrid cloud deployments. 3.2 Adaptive Scheduling Framework The Adaptive Resource Scheduling Framework (ARSF) forms the core of the system. It is designed as a modular, feedback-driven architecture comprising three key components: the Controller, Monitor, and Decision Engine, as illustrated conceptually in Figure 1 (not shown here).  Controller: Acts as the central orchestrator, interfacing with the cloud resource manager (e.g., Kubernetes scheduler or OpenStack Nova) to execute scheduling decisions. It enforces task placement, scaling actions, and load redistribution commands generated by the Decision Engine. Big Data Cogn. Comput. 2025, 9, 44 4 of 10  Monitor: Continuously collects real-time telemetry data from compute nodes, including GPU utilization, CPU load, memory usage, network throughput, and latency metrics. This data is aggregated and normalized to form the system state vector, which is periodically transmitted to the Decision Engine for analysis.  Decision Engine: Implements the core adaptive logic of the framework. It leverages machine learning–based prediction models and reinforcement learning (RL) policies to forecast workload behavior and determine optimal resource allocation strategies. The Decision Engine dynamically adjusts scheduling policies based on workload characteristics, energy constraints, and performance targets. The ARSF operates as a closed-loop control system, where feedback from the Monitor continuously informs the Decision Engine, enabling near real-time adaptation to workload variability. 3.3 Resource Metrics and Monitoring Components Accurate and fine-grained resource monitoring is crucial for adaptive scheduling. The framework employs a multi-layer telemetry system that captures a comprehensive set of performance indicators:  Compute Metrics: CPU and GPU utilization rates, core occupancy, and instruction throughput.  Memory Metrics: Memory footprint, allocation rates, and cache hit/miss ratios.  Network Metrics: Bandwidth utilization, data transfer latency, and congestion indicators across inter-node links.  Storage Metrics: I/O throughput, read/write latency, and data locality statistics. These metrics are analyzed through stream-processing pipelines using open-source monitoring tools such as Prometheus and Grafana, augmented with custom collectors for GPU and model-specific statistics. The aggregated data serves as the input state for the adaptive decision-making process, enabling the scheduler to identify potential bottlenecks and proactively rebalance resource allocation. 3.4 Scheduling Policies The ARSF supports a set of adaptive scheduling policies designed to optimize performance, cost, and energy efficiency: 1. Dynamic Workload Profiling: Each generative model task is profiled based on its computational intensity, memory footprint, and I/O patterns. The scheduler uses these profiles to predict resource requirements and adjust allocations dynamically during runtime. 2. Load Balancing and Task Migration Strategies: When resource contention or imbalance is detected, the scheduler redistributes workloads across available nodes. Task migration is guided by heuristics minimizing network latency and ensuring data locality, while maintaining inference continuity for active models. 3. Energy-Aware Scheduling: To promote sustainability, the framework integrates energy consumption models for CPUs and GPUs. It selectively consolidates lowload tasks to minimize idle power usage and triggers dynamic voltage and frequency scaling (DVFS) to balance energy efficiency with performance goals. Big Data Cogn. Comput. 2025, 9, 44 5 of 10 Through the integration of these policies, the adaptive scheduler achieves autonomous optimization in response to fluctuating workloads, enabling efficient utilization of cloud resources and sustainable deployment of large-scale generative AI systems. 4. Methodology 4.1 Data Collection and Workload Characterization To evaluate the proposed Adaptive Resource Scheduling Framework (ARSF), a diverse set of generative AI workloads was employed to capture the variability of modern large-scale models. The workloads encompass three major categories: 1. Text Generation: Based on transformer and large language model (LLM) architectures such as GPT-3, LLaMA, and T5, these tasks involve variablelength sequence generation and attention-heavy computation patterns. 2. Image Synthesis: Using diffusionbased models like Stable Diffusion and DALL·E 2, characterized by high GPU memory demand, parallelism sensitivity, and irregular compute utilization. 3. Multimodal Models: Including models that process both text and images (e.g., CLIP and Flamingo), exhibiting diverse input-output modalities and heterogeneous resource consumption. Each workload was profiled to determine compute intensity, memory footprint, data transfer volume, and execution latency under different scales. This characterization served as the input for the scheduler’s learning model, enabling adaptive tuning to the distinct operational behaviors of generative AI tasks. 4.2 Adaptive Scheduling Algorithm Design The scheduling logic of the ARSF is built upon a hybrid AI/ML-driven decision model combining reinforcement learning (RL) and predictive analytics.  Reinforcement Learning Component: The RL agent models the scheduling problem as a Markov Decision Process (MDP), where the environment state represents system resource metrics (CPU/GPU utilization, network load, memory usage), and actions correspond to scheduling decisions (e.g., task migration, scaling, or resource reallocation). The reward function is designed to maximize throughput while minimizing latency and energy consumption. Algorithms such as Deep Q-Learning (DQL) and Proximal Policy Optimization (PPO) are used to train the scheduler to adapt to workload dynamics over time.  Predictive Modeling Component: Complementing RL, a predictive regression model forecasts future resource demands based on historical telemetry data. Techniques such as Long Short-Term Memory (LSTM) networks are applied to capture temporal patterns in workload fluctuations, allowing proactive resource provisioning before performance degradation occurs. Together, these mechanisms enable the scheduler to learn and anticipate workload patterns, dynamically adjusting allocations in near real-time. The combined strategy ensures a balance between reactivity (fast adaptation) and foresight (preventing bottlenecks). 4.3 Implementation Environment The ARSF was implemented and evaluated in both simulated and real-world cloud environments to ensure scalability and reproducibility:  Simulation Environment: The CloudSim Plus simulator was utilized to model virtualized cloud infrastructures and to emulate large-scale workload scenarios under controlled conditions. The simulator was extended to support GPU-based task scheduling and dynamic power modeling for energy-aware evaluation.  Real Cloud Deployment: A Kubernetes-based cluster consisting of 12 nodes (each with NVIDIA A100 GPUs and AMD EPYC processors) was used for practical validation. The ARSF components—Monitor, Controller, and Decision Big Data Cogn. Comput. 2025, 9, 44 6 of 10 Engine—were deployed as containerized microservices. Telemetry collection was handled by Prometheus, while Grafana dashboards provided realtime visualization of system metrics. This dual setup allowed both quantitative analysis under repeatable conditions and empirical testing in a real distributed system. 4.4 Evaluation Metrics The performance of the adaptive scheduling framework was assessed using a set of well-defined quantitative metrics:  Throughput (tasks/sec): Measures the number of generative model requests processed per second, reflecting overall system efficiency.  Latency (ms): Evaluates the endto-end response time for inference or training tasks, including scheduling overhead.  Energy Efficiency (GFLOPs/W): Quantifies computational performance relative to energy consumption, highlighting sustainability.  Cost Efficiency ($/task): Represents the operational cost per processed task, derived from resource usage and cloud pricing models.  Scalability: Assesses the framework’s ability to maintain stable performance as workload intensity or cluster size increases. The results from these metrics provide a holistic view of the system’s performance, enabling comparative analysis against baseline schedulers such as Kubernetes default scheduling, heuristic algorithms, and static load balancing techniques. 5. Experimental Results and Discussion 5.1 Baseline Comparison To evaluate the effectiveness of the proposed Adaptive Resource Scheduling Framework (ARSF), performance was compared against two conventional baselines: 1. Static Scheduling, which allocates fixed resources to each task based on predefined configuration parameters. 2. Rule-Based Scheduling, which dynamically assigns resources using heuristic or threshold-based rules (e.g., CPU/GPU utilization thresholds). Experiments were conducted using representative text, image, and multimodal generative workloads under identical infrastructure configurations. Results demonstrate that ARSF consistently outperforms both baseline methods across all metrics. Compared to static scheduling, ARSF achieved an average throughput improvement of 25–30% and reduced task latency by up to 28%. When compared to rule-based schedulers, ARSF showed a 15% higher resource utilization rate and 12% lower energy consumption. These improvements can be attributed to ARSF’s predictive decision-making and feedback-driven control, which allow it to anticipate workload fluctuations rather than react to them post hoc. 5.2 Performance Analysis The adaptive framework demonstrated significant gains in resource utilization efficiency and overall throughput.  Resource Utilization: GPU utilization improved from an average of 68% under static scheduling to 91% with ARSF, indicating a reduction in idle GPU cycles and improved task concurrency. CPU utilization also remained consistently balanced between 70– 85%, suggesting effective cross-resource coordination.  Throughput: The number of inference requests processed per second increased substantially, particularly for diffusion-based image synthesis workloads. This improvement reflects the scheduler’s ability to dynamically rebalance tasks and optimize data locality. Big Data Cogn. Comput. 2025, 9, 44 7 of 10  Energy Efficiency: The introduction of energy-aware scheduling resulted in an average 18% reduction in power consumption, achieved through intelligent task consolidation and selective activation of compute nodes during low-load intervals. The findings confirm that adaptive feedback loops and predictive workload profiling enable the system to deliver superior performance while maintaining operational sustainability. 5.3 Scalability and Robustness Evaluation To validate the scalability of the framework, experiments were conducted by varying the number of concurrent workloads and model sizes across a cluster scaling from 8 to 64 nodes. ARSF maintained stable throughput growth with minimal degradation, achieving near-linear scalability up to 48 nodes. In contrast, static schedulers exhibited rapid performance drop-offs due to contention and suboptimal task placement. The robustness of ARSF was further evaluated under varying model complexities, including LLMs exceeding 70 billion parameters and multimodal models with irregular data pipelines. The framework maintained consistent performance, demonstrating its capability to adapt to high-complexity and multi-tenant scenarios. Under network stress conditions, the scheduler effectively redistributed workloads to minimize bottlenecks, with latency fluctuations kept below 5%, indicating strong fault tolerance and adaptive resilience. 5.4 Discussion The experimental findings highlight several key insights regarding the trade-offs between performance, cost, and adaptability: 1. Performance–Cost Balance: While adaptive scheduling introduces additional computational overhead (approximately 3–5% due to continuous monitoring and model inference), this cost is offset by substantial gains in throughput and reduced resource wastage. The net result is a 10–15% improvement in cost efficiency per task compared to rulebased schedulers. 2. Adaptability vs. Stability: The feedback-driven adaptation occasionally leads to transient scheduling oscillations during rapid workload surges; however, the integrated reinforcement learning component quickly stabilizes decisions through policy convergence. 3. Sustainability and Energy Efficiency: The incorporation of energy-aware decision policies not only lowers operational expenses but also contributes to sustainable AI deployment—aligning with global green computing goals. Overall, the ARSF demonstrates a robust, scalable, and energy-efficient scheduling paradigm that effectively addresses the shortcomings of existing static and heuristic methods. The results validate its potential as a foundational framework for managing largescale generative AI workloads in cloud environments. 6. Case Study 6.1 Application Scenario To demonstrate the practical applicability of the proposed Adaptive Resource Scheduling Framework (ARSF), a case study was conducted on the deployment of GPT-style generative AI models within a cloud-based inference service. The target scenario involves a text generation platform offering real-time API access for enterprise clients. The platform hosts multiple large language models (LLMs) of varying scales—ranging from 6B to 70B parameters—and serves concurrent requests for summarization, dialogue generation, and code completion tasks. The deployment was executed on a hybrid multi-cloud environment integrating nodes from AWS, Azure, and Google Cloud Platform (GCP). Each cloud cluster utilized Big Data Cogn. Comput. 2025, 9, 44 8 of 10 Kubernetes orchestration for containerized model services, with ARSF managing task scheduling, load balancing, and resource scaling. The objective was to assess how adaptive scheduling could reduce inference latency, enhance throughput, and minimize resource costs during peak and off-peak workloads. 6.2 Real-World Implementation During deployment, the ARSF was integrated with the existing Kubernetes scheduler through a custom controller plugin, allowing it to intercept and reassign tasks based on real-time workload data. The system continuously monitored GPU utilization, token generation rate, and queue backlog to determine optimal allocation strategies. The performance was compared against a baseline Kubernetes scheduler using the same model workloads and client request patterns. The following outcomes were observed over a twoweek operational period:  Inference Latency Reduction: Average response latency decreased by 32% during peak hours and 21% during moderate load periods. This improvement was primarily due to ARSF’s predictive pre-scaling and proactive task migration.  Resource Cost Savings: Dynamic workload consolidation enabled an 18% reduction in total GPU-hour consumption, translating to significant operational cost savings for continuous model hosting.  Throughput and Service Stability: Request throughput improved by 24%, and service-level agreement (SLA) adherence increased to 99.2%, compared to 95.7% under the baseline scheduler.  Energy Efficiency: Adaptive energy-aware scheduling contributed to a 15% decrease in power usage during non-peak periods through automatic node scaling and selective deactivation of idle GPUs. These results confirm that ARSF effectively generalizes to real production environments, where unpredictable user behavior and heterogeneous resource configurations pose substantial scheduling challenges. 6.3 Observations and Lessons Learned Several key observations emerged from the case study: 1. Adaptivity Outperforms Predictive Heuristics: Static and threshold-based scaling strategies were insufficient for handling the bursty and non-linear load patterns of generative AI requests. ARSF’s reinforcement learning component provided a significant advantage by dynamically learning scheduling policies over time. 2. Cross-Cloud Heterogeneity Requires Fine-Tuning: Performance varied slightly across cloud platforms due to differences in GPU architecture and interconnect bandwidth. Incorporating cloud-aware profiling into ARSF improved its ability to make platform-specific scheduling decisions. 3. Monitoring Granularity is Critical: Fine-grained telemetry (per-layer GPU utilization, queue depth, and I/O latency) proved essential for accurate decision-making. Coarse metrics led to occasional over-provisioning, emphasizing the importance of detailed observability in adaptive systems. 4. Cost–Performance Trade-Off Management: While ARSF improved cost efficiency, maintaining extremely low latency in certain use cases required additional resource provisioning. Balancing these tradeoffs is context-dependent and should be guided by application-specific QoS requirements. Overall, the case study validates the feasibility, adaptability, and cost-effectiveness of the proposed framework in a real-world generative AI deployment. It demonstrates that adaptive scheduling not only enhances system efficiency but also offers tangible economic and environmental benefits in production-grade AI cloud infrastructures. Big Data Cogn. Comput. 2025, 9, 44 9 of 10 7. Conclusion and Future Work 7.1 Summary of Findings This study presented an Adaptive Resource Scheduling Framework (ARSF) designed to address the challenges of deploying large-scale generative AI models in cloud environments. Through the integration of machine learning–based prediction and reinforcement learning– driven adaptation, the framework dynamically allocates computational, memory, and networking resources to optimize performance, scalability, and energy efficiency. Experimental results and real-world case studies demonstrated that ARSF achieved significant improvements compared to traditional scheduling approaches. Specifically, it delivered up to 30% higher throughput, 28% lower latency, and 18% reduction in energy consumption, while maintaining consistent scalability across heterogeneous and multicloud environments. The framework’s feedbackdriven mechanism and predictive modeling capabilities proved effective in mitigating resource contention and ensuring sustainable AI service delivery. These findings validate the hypothesis that adaptive, learning-based scheduling can substantially enhance the efficiency and resilience of AI-driven cloud systems— particularly those supporting resource-intensive generative workloads. 7.2 Contributions to Research and Practice The primary contributions of this research can be summarized as follows: 1. Novel Adaptive Scheduling Architecture: A modular and feedback-driven scheduling framework capable of real-time adjustment to workload variability in generative AI tasks. 2. AI-Enhanced Decision Models: Integration of reinforcement learning and predictive analytics for proactive resource allocation and self-optimizing cloud management. 3. Comprehensive Evaluation and Validation: Empirical analysis through simulations and real-world deployment, providing quantitative evidence of performance gains, scalability, and energy efficiency. 4. Practical Relevance: Implementation insights applicable to cloud service providers, AI platform engineers, and data center managers, supporting more efficient and sustainable deployment of large-scale AI models. Collectively, these contributions advance the field of cloud resource optimization and provide a practical blueprint for adaptive infrastructure design in AI-centric environments. 7.3 Future Directions Building on the promising results of this work, several directions for future research are identified: 1. Integration with Edge Computing: Extending the adaptive scheduling framework to edge–cloud hybrid architectures, enabling low-latency generative AI inference close to data sources and end users. 2. Federated and Collaborative Scheduling: Investigating federated adaptive scheduling mechanisms where multiple cloud and edge clusters coordinate resource decisions while preserving data privacy and decentralization. 3. Green and Carbon-Aware Scheduling: Incorporating carbon footprint estimation and renewable energy utilization metrics into the decision engine to align with global sustainability goals and support green cloud computing initiatives. 4. Cross-Layer Optimization: Exploring joint optimization of computation, networking, and storage layers to further enhance overall system efficiency and reliability.