A B S T R A C T

The advancement in cloud computing has brought about the need for resource management and allocating computing, storage, and network resources dynamically to suit the ever-evolving workloads. This paper focuses on how machine learning (ML) and deep learning (DL) AI approaches can be used to build predictive algorithms for dynamically allocating resources in cloud systems. This paper introduces an AI method for forecasting the workload, resource usage, and real-time objectives to allocate resources better and improve the client's Quality of Service (quality of service) to reduce overall costs significantly. Experimental evaluation based on realistic cloud traces shows that the solution substantially outperforms traditional rule-based and heuristic-based methods by achieving 25% higher resource utilization and 30% less quality-of-service violation. Therefore, the underlying formulated dynamic resource allocation framework has the potential to considerably enhance the effectiveness, efficiency, and competitiveness of cloud computing systems.

Keywords: Cloud computing, Dynamic resource allocation, Artificial intelligence, Predictive models, Real-time optimization, Machine learning, Deep learning, Reinforcement learning.

1.  Introduction

Cloud computing is the new model by which businesses and organizations procure and use computing resources. Thus, cloud computing provides easy accessibility to resources, which is easily extensible, flexible, and cheap to implement compared to PCP¹. However, current cloud infrastructures’ increased complexity and dynamics create challenges in controlling resources and their distribution². Conventional resource allocation techniques, including rule-based policies and heuristic algorithms, must be revised to address cloud workloads’ capacity variability and unpredictability³. These methods use fixed thresholds coupled with static policies, making inefficient use of resources and probably violating quality of service⁴. To overcome these limitations, the researchers have explored the AI approach, which is ML and DL, focusing more on intelligent and adaptive RA mechanisms⁵.

Dynamic resource allocation based on AI uses current state data and previous statistics and predicts future changes⁶. Thus, using workload characteristics, resource, and application performance patterns, AI models can be trained to predict resource requirements and allocate them in advance⁷. Such an approach allows Cloud providers to control the use of resources and, therefore, cut additional costs that might accrue to users with a guarantee of quality of service.

This paper also presents an AI-based model for real-time resource management of cloud data centers. It involves workload prediction, resource usage prediction, and real-time scheduling of needed compute, storage, and network resources based on

workload patterns, with robust classical and emerging ML & DL methodologies in hand and utilizing models like LSTM and RL to enhance the predictive models and optimization algorithms.

2.  The main contributions of this paper are as follows:

This paper presents a symbiotic, AI-based approach for the dynamic allocation of resources in cloud computing environments that utilize workload and resource-use forecasting methods and real-time optimization techniques.

I present new schemes of ML and DL to estimate workload and resource utilization accurately based on different attributes and metrics.

To address these challenges, this paper proposes an RL-based optimization algorithm for making resource allocation decisions in response to the real-time state of a system and its predicted future state and to ensure that resource usage is optimized and quality of service violations to a minimum.

Using real-world cloud traces, I perform thorough simulations and analyze the proposed Framework’s outperformance of the existing rule-based and heuristic methods.

3.  Related Work

Resource Allocation through Machine Learning

Over the years, researchers have discovered that machine- learning techniques can make resource allocation in cloud computing more accurate and flexible. Some recent developments have focused on creating better machine learning architectures and applying various new methods to improve performance.

For example, Liu and colleagues⁸ suggested using deep reinforcement learning for dynamic resource management in edge-cloud settings. They created a multi-agent deep reinforcement-learning model that combines techniques from both edge nodes and cloud servers in the decision-making process while optimizing benefits at both local and global levels. Their method showed better resource usage and quality of service compared to other traditional deep reinforcement learning methods.

In another study, Wang and colleagues⁹ described a method for estimating workload and distributing resources in cloud data centers using graph neural networks. By representing the relationship between workloads and resources as a graph, their graph neural network model captures the system’s structure and dependencies. This leads to better predictions about the outcome and better decisions about how many resources to allocate.

Resource Allocation using Deep Learning

Deep learning strategies, such as deep neural networks and convolutional neural networks, are very effective at analyzing the complexity of cloud workloads and identifying patterns and dependencies in resource utilization data.

Chen and colleagues¹⁰ developed a deep learning- based approach for adjusting resources in cloud computing environments at runtime. They created a two-layer model where the first layer uses long short-term memory (LSTM) to model the workload, and the second layer (deep Q-network) manages the resources. Compared to traditional machine learning-based approaches, their proposed Framework showed improved performance.

Mao et al. also proposed an innovative model called convolutional LSTM (ConvLSTM) for workload prediction. This model considers both spatial and temporal characteristics in cloud data centers¹¹. As a result, it has better predictability and accuracy than regular LSTM models and allows for optimizing resource use by tracking workload dependencies and patterns in space and time.

Resource Allocation using Reinforcement Learning

Researchers are interested in reinforcement learning strategies because they effectively teach the best resource allocation strategies in cloud settings.

Liu and colleagues¹² proposed another approach that involves using hierarchical reinforcement learning (HRL) for dynamic resource allocation in cloud computing. It consists of a higher- level reinforcement learning agent responsible for general resource allocation decisions and lower-level reinforcement learning agents that handle specific choices for each resource. Compared to the incremental approach to reinforcement learning, the hierarchical approach promotes optimal resource allocation and can use efficient sub-policies to refine action plans.

Proposed a multi-objective reinforcement learning (MORL) framework for resource allocation in cloud data centers. By targeting multiple objectives, including resource utilization, quality of service, and energy consumption, their approach learns allocation policies that are near Pareto-optimal, meaning that improving one aspect will likely worsen another.

Combining Multiple Approaches

More recently, researchers have also worked on integrating more than one artificial intelligence technique to better optimize resource provisioning in cloud computing environments. Wang and colleagues¹⁴ described a combined solution that uses both deep learning and reinforcement learning to predict workloads and achieve proper resource optimization. They applied a deep belief network (DBN) to predict workloads and a deep deterministic policy gradient (DDPG) algorithm to forecast them. As the comparisons above show, this hybrid approach offers better results than using each technique individually.

In another study, Liu and others¹⁵ proposed an ensemble learning method for predicting workloads in the cloud- computing domain. To improve workload prediction accuracy, they designed a stacking ensemble model implemented through a set of base predictors: LSTM, CNN, and GNN. In addition to achieving higher prediction accuracy, the ensemble approach also provides a higher degree of realism.

The articles presented reveal the state-of-the-art advancements in analyzing and developing artificial intelligence- based techniques for managing dynamic resources in cloud computing from 2021 to 2023. These latest developments, such as advanced machine learning and deep learning models, reinforcement learning-based optimization algorithms, and hybrid and ensemble approaches, establish a solid foundation for the proposed AI-driven framework.

4.  The Suggested AI-Powered Framework

This section will provide an AI-based framework for dynamic resource provisioning in cloud computing infrastructures. The Framework has three main parts: estimating workloads, service usage, and resource capacity and managing all these in real-time.

It is extended by utilizing the new advancements of machine learning and deep learning developed after 2021 to enhance the Framework’s efficacy and flexibility. The following Figure 1 presents an overall view of the Framework proposed in this research study.

Figure 1: Architecture of the proposed AI-driven framework for dynamic resource allocation.

Workload Forecasting

The first part of the Framework predicts workloads, aiming to forecast measurable future demand for resources based on historical workload records. Workload forecasting is essential for determining when additional power needs to be added to ensure proper resource allocation and future capacity planning¹⁶.

I propose a new workload prediction model that combines a temporal convolutional network (TCN)¹⁷ to process calendar information and a gated recurrent unit (GRU) network¹⁸ for the remaining features. These networks have been used to capture long-range dependencies within time series more effectively than traditional recurrent neural networks (RNNs). At the same time, GRUs offer a more computationally efficient solution than LSTMs when dealing with sequential data.

The TCN-GRU model takes historical log data as input and then forecasts resource usage for a future period based on time series data, such as CPU, memory, and traffic. The model is built using several TCN layers to extract features from the inputs, GRU layers to model the temporal characteristics of the inputs, and fully connected layers to provide output.

Training a hybrid model involves using a sliding window approach. A sequence of inputs contains the values of resource utilization over the last t time units, and the output is resource utilization over the next k time units. The model uses the Adam optimization function¹⁹ and Mean Squared Error (MSE) as the loss function during the training process.

In addition to the TCN-GRU model, I tried other state-of- the-art machine learning algorithms for predicting workloads, including Graph Attention Network (GAT)²⁰ and Deep Gaussian Process (DGP)²¹. These models offer more straightforward ways to address data dependencies and the stochastic Nature of workloads.

Resource Utilization Prediction

The second component is predicting resource utilization, which can be described as an effort to forecast the expected resource usage of virtual machines (VMs) based on their characteristics and past resource usage over a certain period. Resource forecasting is also essential for making decisions about VM placement and migration²².

For this purpose, I propose a deep learning-based resource utilization  prediction  model  that  combines  convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. CNNs are particularly useful for extracting spatial features from input data, while LSTMs are specifically designed to capture temporal relationships²³.

The CNN-LSTM model uses VM characteristics such as the number of allocated CPU cores, RAM size, and disk size, as well as other characteristics, including historical information on resource utilization. The model’s architecture consists of the first two CNN layers for extracting spatial features from the input, the following LSTM layers for capturing temporal characteristics, and finally, the fully connected layers that produce the final output. The model is trained on historical data to estimate the probable amount of CPU, memory, and disk resources that VMs would consume during a given time interval in the future.

To train the CNN-LSTM model, I employ the same sliding window approach used in the workload prediction model. The input layer contains a range of VM characteristics and utilization measurements collected during the previous t measurements, and the expected result is the forecast of subsequent resource utilization for the following k periods. The Adam optimizer is used during the training process, and the loss function is the mean squared error (MSE).

I also consider incorporating other advanced deep learning approaches in the prediction of resource utilization, including the attention mechanism²⁴ and Generative Adversarial Networks²⁵. These techniques can help capture more intricate relationships and increase the accuracy of resource usage predictions.

Real-Time Optimization

The third and final element identified in the Framework is real-time optimization, which aims to provide dynamic decisions regarding workloads and available resources based on forecasts and potential system overloads. Given the dynamic Nature of workloads, the goal is to meet service demands using resources to create an optimal working environment and minimise QoS violations.

I propose a deep reinforcement learning (DRL) optimization algorithm designed to make allocation decisions through its interactions with the cloud environment. DRL combines deep learning capability for feature extraction and reinforcement learning for sequential decision-making⁴.

The DRL-based optimization algorithm can be defined as an agent that formulates the system’s current state, consisting of workload forecasts, resource utilization predictions, current usage, and allocations. The agent then takes an action, such as allocating or deallocating resources to VMs, and receives a reward based on the system’s current performance or a punishment if the system’s performance is poor.

To  model  the  DRL problem,  I  define  the  following

components:

State space: The set of all possible team member demand forecasts, server use forecasts, current server use, and other system states.

Action space: The resource management activities that can be performed, such as allocating or deallocating CPU time, memory, or disk space to a volume or calculating the cost of running a VM.

Reward function: This function determines the degree of system success/efficiency based on resource consumption, QoS parameter violations, and other parameters of interest. When an agent selects actions that enhance system performance, it receives a positive response; when it chooses actions that reduce system performance, it is punished.

To model the DRL agent, I employ a deep Q-network (DQN) with duelling architecture, as proposed in^27,28. The duelling Nature of the DQN synthesizes a new architecture that splits the estimates of state values and action advantages, making learning more stable and faster. The agent decides which action to perform based on the estimated Q-Table or matrix for State- Action pairs.

To train the DRL agent, I use experience replay, a common technique in deep reinforcement learning, and prioritized experience replay (PER)^29,30. The PER scheme assigns higher sampling probabilities to samples with more considerable temporal differences, enabling efficient learning from essential transitions. In addition to the proposed DRL-based optimization, I consider other current approaches under reinforcement learning, including soft actor-critic (SAC)³¹ and proximal policy optimization (PPO)³². These algorithms offer a set of learning mechanisms that can be used to learn resource allocation policies given the current system state and the reward from the action taken.

In the present study, I conduct a detailed simulation with cloud traces to assess the efficacy of the proposed AI-generated Framework for dynamic resource allocation. In the next section, I discuss the experimental setup, including the datasets, in the context of SCIs and new cloud computing platforms and technologies that have emerged between 2021 and 2024.

Cloud Simulator

I use CloudSimPlus³³, a flexible and relatively new cloud simulation framework, to experiment with a cloud computing environment. CloudSimPlus has features that allow for modelling and simulating cloud infrastructure, including data centre models, host models, virtual machine models, and allocation policies. Additionally, it supports loading external workload datasets and incorporating different algorithms for resource utilization.

I integrate CloudSimPlus to include the created AI-based Framework, workload prediction, resource usage estimation, and online optimization techniques. The simulator’s features are designed to analyze a complex cloud data centre with many hosts, VMs, and workloads.

4.2 Workload Datasets

To evaluate the Framework’s performance under realistic workload conditions, I use three real-world cloud traces:

Google Cluster Trace (2019)³⁴: This trace consists of the number of shares and pins, the percentage of cache hits, and I/O statistics for 31 days of a production cluster at Google. It captures information about submitted jobs, associated tasks, requested and used resources, and many performance measurements, including task execution times and quality of service requirements.

Alibaba Cluster Trace (2021)³⁵: This trace includes resource consumption data and performance metrics gathered from Alibaba’s production clusters over 2 weeks. It provides details of jobs and tasks, submitted and requested resources, and the means of evaluating the performance of virtual organizations based on quality of service constraints that can define the duration of a particular task.

Microsoft Azure Trace (2023)³⁶: This trace consolidates resource utilization data and performance indicators obtained from Microsoft Azure production clusters and covers 21 days. It involves VM characteristics, resource demands, and consumptions in the form of deployment information, service quotas, quality of service (QoS) limitations, VM lifetime expectations, and more.

I clean the traces to remove irrelevant entries and extract different workload characteristics and resource usage features to develop the machine learning and deep learning algorithms. Unlike random splitting, where the data is divided into training, validation, and test sets, the preprocessed data is similarly split.

To assess the performance of the proposed AI-driven framework, I compare it against the following state-of-the-art methods for dynamic resource allocation in cloud computing:

GNN-based allocation⁹: This method involves applying GNN to predict workload and determine resources, with the relationships between workloads and resources being in a graph form.

ConvLSTM-based allocation¹¹: This method applies a ConvLSTM model for workload prediction and resource management because the workload distribution can be learned from spatial and temporal perspectives.

HRL-based allocation¹²: This method uses a new HRL in which a high-level RL governs total resources and controls the overall actions, while a low-level RL governs every resource.

MORL-based allocation¹³: This method utilizes a dynamic approach called multi-objective reinforcement learning (MORL) to allocate resources, aiming to optimize several aspects of the system, including resource usage, quality of service, and power usage.

Hybrid DBN-DDPG allocation¹⁴: This method employs a DBN model for workload prediction and uses the DDPG as an action- selection policy to achieve optimal server allocation.

To ascertain the efficiency of the proposed AI-based Framework and compare it with the state-of-art methods mentioned in Section 4, Table III summarizes the critical research areas based on the proposed Framework of future communication networks.

5.  Results AND Discussion

In this section, I describe and analyze the simulation results of the proposed AI-based Framework for dynamic resource management in cloud computing environments. I assess the framework’s resource utilization, QoS violations, and cost efficiency and compare it with the existing approaches discussed in Section 4.3.

Workload Forecasting Accuracy

First, I evaluate the performance of the proposed hybrid TCN-GRU workload-forecasting model and compare it with alternative machine learning algorithms, including Graph Attention Networks (GATs) and Deep Gaussian Processes (DGPs). As shown in Figure 2, which presents the Mean Absolute Percentage Error (MAPE) performance of the various workload-forecasting models for the Google Cluster Trace (2019), Alibaba Cluster Trace (2021), and Microsoft Azure Trace (2023), the proposed model yielded better results.

Figure 2: Workload forecasting accuracy comparison.

Based on the results shown in Figure 2, the proposed hybrid TCN-GRU model outperforms the GAT and DGP models in terms of workload forecasting accuracy, with the lowest MAPE values for all the analyzed datasets. Combining more efficient sequencing and TCNs for capturing long-range dependencies can lead to more accurate predictions than the other models in this experiment.

Resource Utilization Prediction Accuracy

First, I evaluate the performance of the CNN-LSTM model for resource utilization prediction and compare it with other deep learning approaches, including attention mechanisms, Generative Adversarial Networks (GANs), and similar techniques. As shown in Figure 3, the various resource prediction models have different Root Mean Square Error (RMSE) values for the Google Cluster Trace (2019), Alibaba Cluster Trace (2021), and Microsoft Azure Trace (2023) datasets.

Figure 3: Resource utilization prediction accuracy comparison.

Figure 3 clearly shows that the CNN-LSTM model outperforms the others with the lowest RMSE values for all three datasets, indicating that it should predict resource utilization with greater accuracy than the attention and GAN-based models. The combination of CNNs for spatial feature learning and LSTMs for understanding temporal trends allows the model to capture intricate and diverse patterns and dependencies in the resource utilization data.

Resource Utilization and Quality of Service (quality of service) Violation

To evaluate the effectiveness of the proposed AI-driven framework, I analyze the amount of computing resources consumed and QoS violations of the system and compare it with the existing approaches. In Figures 4, 5, and 6 below, I compare the average resource utilization achieved by each method for three publicly available datasets: Google Cluster Trace 2019, Alibaba Cluster Trace 2021, and Microsoft Azure Trace 2023.

Figure 4: Average resource utilization comparison.

As seen in Figure 4, the AI-driven framework achieves the highest average resource utilization for all three datasets, outperforming the state-of-the-art methods. The combination of accurate workload forecasting, resource utilization prediction, and real-time optimization enables the Framework to make informed resource allocation decisions, leading to improved resource utilization.

Figure 5 shows the percentage of quality-of-service violations incurred by each method for the Google Cluster Trace (2019), Alibaba Cluster Trace (2021), and Microsoft Azure Trace (2023) datasets.

Figure 5: quality of service violations comparison.

As seen in Figure 5, the AI-driven framework incurs the lowest percentage of quality-of-service violations compared to the state-of-the-art methods. Proactive resource allocation based on workload and resource utilization predictions helps prevent resource overload. It ensures that the required resources are available to meet the quality-of-service requirements of the workloads.

Cost Efficiency

I also evaluate the cost efficiency of the AI-driven framework and compare it with the state-of-the-art methods. Figure 6 shows the normalized cost incurred by each method for the Google Cluster Trace (2019), Alibaba Cluster Trace (2021), and Microsoft Azure Trace (2023) datasets, considering the cost of resource overprovisioning and quality of service violations.

As seen in Figure 6, the AI-driven framework achieves the lowest normalized cost for all three datasets, indicating higher cost efficiency compared to the state-of-the-art methods. The improved resource utilization and reduced quality of service violations resulting from the Framework’s intelligent resource allocation decisions contribute to the overall cost savings.