Machine Learning (ML)-Centric Resource Management in Cloud Computing: A Review and Future Directions

1. Introduction

• Variable and unpredictable VM workloads can overutilize or underutilize servers, causing inconsistent QoS, SLA violations, imbalanced energy use, and higher costs and carbon footprints.Reported average CPU and memory utilization was 17.76% and 77.93%, while a Google cluster could not exceed 60% and 50%, respectively; data-centre energy use was around 35 Twh in 2015 and expected to reach 95 Twh by 2040.Source: §1 · p. 3

1.2. Our Contributions

• The review covers ML-based resource-management challenges, methods, advantages, limitations, experimental configurations, datasets, performance improvements, and proposed future research directions.Its scope includes provisioning, VM consolidation, thermal prediction, and other management approaches.Source: §1 · p. 4, §1.2 · p. 6

1.1. Motivation of Research

• ML-based resource management addresses nonlinear workload dynamics by forecasting demand and enabling dynamic provisioning, VM consolidation, scheduling, and thermal management while improving QoS and energy efficiency.Static threshold policies can impose overhead, delay responses, and trigger unnecessary VM migration; ML replaces them with adaptive, data-driven heuristics [11].Source: §1 · p. 4, §1.1 · p. 5

1.3. Related Surveys

• Prior surveys address resource usage, provisioning, allocation, mapping, adaptation, migration, forecasting, stability, and availability, but provide limited detail on ML-specific challenges and future directions.This article therefore presents a broader survey of ML algorithms, shortcomings, challenges, and potential research directions in data-centre resource management.Source: §1.3 · pp. 6–7

1.4. Article Structure

• The paper proceeds from background and definitions to challenges and limitations, future research directions, and a concluding section.These topics are covered in Sections 2 through 5, respectively.Source: §1.4 · p. 7

2.1. Cloud Computing

• Cloud computing provisions resources such as memory, CPU, bandwidth, disc, and applications or services over the Internet.Source: §2.1 · p. 7
• NIST defines cloud computing as on-demand network access to a shared pool of configurable resources that can be rapidly provisioned and released with minimal management effort.Source: §2.1 · p. 7
• The NIST cloud model includes five core features, three service models, four deployment options, and two additional characteristics identified in the literature.Source: §2.1 · p. 7
• Cloud computing uses a client-server architecture to support centralized application deployment and computation offloading.Source: §2.1 · p. 7
• The model enables cost-effective application delivery and maintenance while allowing flexible resource provisioning and detachment from related technologies.Source: §2.1 · p. 7
• Commercial cloud systems include Alibaba Cloud, Microsoft Azure, Adobe Creative Cloud, ServerSpace, Amazon Web Services, and Oracle Cloud.Source: §2.1 · p. 7

Core features

• On-demand self-service lets clients request services as needed and pay through an online “pay-and-go” control center without human interaction.Source: §2.2 · p. 7
• Broad network access enables standardized access to resources across provider areas from diverse locations and incompatible thin or thick clients.Source: §2.2 · p. 8
• Resource pooling presents diverse real or virtual services as one location-agnostic blended resource, allowing vendors to include resources dynamically.Source: §2.2 · p. 8
• Rapid elasticity allows clients to scale resources up or down and request services whenever needed, exemplified by Amazon Elastic Compute Cloud.Source: §2.2 · p. 8
• Measured service requires automatic control, monitoring, optimization, and documentation, while multi-tenancy and auditability require policy-based governance, billing, logs, and compliance assessment.Source: §2.2 · p. 8

• SaaS lets clients access provider-hosted applications through web portals, while provider access simplifies production and testing.Source: §2.3 · p. 8
• PaaS provides network, servers, and an operating system so clients can build applications and manage their configuration settings.Source: §2.3 · p. 9
• IaaS serves users who have created their applications and need basic infrastructure, with vendors potentially providing processors, networks, and storage.Source: §2.3 · p. 9

Public cloud

• Public clouds provide Internet-based services under predetermined rules and offer diverse resource choices while maintaining QoS.They are described as the most popular deployment model and rely on substantial shared resource bases.Source: §2.4 · p. 9

Private cloud

• Private clouds offer public-cloud advantages within an organization, improving security through corporate firewalls but imposing high establishment and management costs.The managing business remains responsible for all aspects of the system.Source: §2.4 · p. 9

Community cloud

• Community clouds let organizations share infrastructure under common criteria, concerns, and policies, reducing costs through sharing while providing high protection.Infrastructure may be supplied by a third-party provider or by community members.Source: §2.4 · p. 9

Hybrid cloud

• Hybrid clouds combine independent public, private, or community clouds while preserving constituent features and requiring standardized functions for application and data interoperability and portability.The combined infrastructures must communicate through agreed-upon functionalities.Source: §2.4 · p. 9

Machine Learning

• Machine learning builds prediction algorithms from training data rather than explicitly programming answers for every possible input.Its widespread adoption has supported applications including computer vision, speech recognition, natural language processing, robotics, self-driving cars, web search, and recommendations.Source: §2.5 · p. 10
• Supervised learning learns feature-to-label mappings for classification or regression using labelled samples.It can classify objects from their characteristics or forecast continuous variables such as stock prices.Source: §2.5 · p. 11
• Unsupervised learning learns data distributions without labels, including clustering VMs by resource-use patterns.Its purpose is to show how data points vary from one another and discover groupings.Source: §2.5 · p. 11
• Semi-supervised learning combines labelled and unlabelled information to improve classification or clustering.Unknown-label data can assist classification, while known same-class relationships can support clustering.Source: §2.5 · pp. 11–12
• Reinforcement learning trains agents through environmental interaction, balancing exploration and exploitation with rewards or sanctions rather than labelled input-output pairs.It is commonly used for robot and computer-game agents.Source: §2.5 · p. 12

2.6. Optimization objective in machine learning

• Machine-learning algorithms optimize objective functions, whose formulation precedes applying numerical optimization methods to evaluate their extrema.Source: §2.6 · p. 12

2.6.1. Optimization in Supervised Learning

• Supervised learning minimizes training loss through a mapping function, using losses such as squared Euclidean distance while regularization and cross-validation address overfitting and model selection.Source: §2.6.1 · pp. 12–13

2.6.2. Optimization in Unsupervised Learning

• Unsupervised optimization minimizes within-cluster variance in k-means and reconstruction error in PCA while preserving data information after dimensionality reduction.Source: §2.6.2 · p. 13

2.6.3. Optimization in Reinforcement Learning

• Reinforcement learning optimizes a policy mapping states to actions by maximizing cumulative discounted returns that depend on environmental rewards.Source: §2.6.3 · pp. 13–14

2.6.4. Optimization in Semisupervised Learning

• Semisupervised learning combines labelled and unlabeled data across tasks, with S3VM optimization using penalty and slack variables to represent classification errors.Source: §2.6.4 · p. 14

3.1. Performance and online profiling of workload

• ML-based resource management research must address online workload profiling, accurate prediction, and low time complexity because production behavior may be unavailable beforehand.Offline profiling is often infeasible, while online profiling must determine when a random VM exhibits representative behavior.Source: §3.1 · p. 15
• Microsoft Azure’s prediction system provides online predictions to multiple resource managers and uses RC-benefit forecasts to reduce physical-resource overuse, but omits memory utilization and distributed-data-center labeling challenges.The system supports server health, migration, container scheduling, and energy-capping managers; EGBT classification may lack sufficient labels when data is distributed.Source: §3.1 · p. 15

3.2. Multiple Resource Usage in VM Consolidation

• VM consolidation based only on current CPU utilization can cause unnecessary migrations, host overutilization, and missed memory or disk bottlenecks.Future utilization estimation across multiple resources is needed for more effective consolidation decisions.Source: §3.2 · p. 21
• [76] used Linear Regression with PlanetLab traces and CloudSim to predict VM utilization, reducing energy consumption while accounting for time overhead on a 7600-host simulated benchmark.Training time and feature dependence may make the approach too slow for real-world data-center response requirements.Source: §3.2 · p. 21

3.3. Cloud Network Traffic

• Single-VM allocation and current-utilization placement can waste resources as application demands fluctuate across workloads and hosts.Placement strategies need future resource-demand estimates to accommodate changing CPU requirements and heterogeneous application behavior.Source: §3.3 · pp. 21–22
• Cloud network bandwidth is an increasingly important placement constraint: projected cloud-computing traffic of 51,774GB/sec can increase VM migration time and violate SLAs [91][92][93].Network effects therefore materially constrain dynamic VM placement and resource management.Source: §3.3 · p. 22
• [77] predicted CPU utilization and network bandwidth for dynamic, energy-efficient VM placement that reduces SLA violations, but excluded disk throughput, another migration-time factor.The study shows the value of network-aware prediction while leaving multi-resource migration modeling incomplete.Source: §3.3 · p. 22

3.4. Host Temperature

• Up to 6.5° lower peak host temperature and 34% lower energy consumption were reported by thermal-aware predictive scheduling, although ambient-temperature prediction may increase overhead [78].The approach predicts host temperature and guides VM migration to minimize peak temperature while using fewer hosts.Source: §3.4 · pp. 22–23

3.5. False Host Overloaded Detection

• Multiple current and predicted resource-utilisation steps can reduce unnecessary VM migrations, but consolidation remains unresolved when a host is currently overloaded yet predicted not to remain overloaded [79].The challenge is distinguishing transient overloads from sustained degradation to avoid migration overhead and wasted energy.Source: §3.5 · pp. 23–24

3.6. Energy metering at Software-Level

• Individual-VM energy monitoring is difficult, while server-level energy capping can degrade co-located workloads, breach SLAs, and violate virtualization isolation; iMeter used SVR and PCA to predict VM energy [80].Different CPU, memory, and I/O demands complicate prediction and can slow per-VM resource decisions.Source: §3.6 · p. 24

3.7. SLA-based VM Management

• SLA-aware management combined historical CPU utilisation with SLA penalties and ANN forecasts for compute-intensive and transactional applications, but deviations, low predictions, and untested non-linearity limited reliability [81].The approach supports placement decisions across varied SLA specifications, including under-utilisation during usual or off-peak periods.Source: §3.7 · p. 25

3.8. QoS-Aware Resource Provisioning

• Dynamic provisioning must determine resources under varying demand because static allocation wastes capacity at low load and weakens QoS at high load; ARIMA-based provisioning used forecast requests, response time, and rejection rate [82].A fixed forecasting interval may harm response time when VM deployment finishes before that interval ends.Source: §3.8 · p. 26

3.9. Varying Patterns of a Service Tenant in Resource Allocation

• Multi-tenant demand prediction requires continual model updates because tenant requirements vary with operation type, timing, and load, complicating proactive provisioning for one or many tenants.Changing tenant profiles and resource requirements are central challenges in service-cloud resource allocation.Source: §3.9 · pp. 26–27
• The proposed tenant-grouping approach predicted only tenants expected to increase demand and used Best-fit decreasing for PM allocation, reducing prediction cost but relying on costly or unavailable binary labels.Unlabeled data in large distributed clouds can make supervised classification unusable.Source: §3.9 · p. 27

3.10. Single ML model in energy consumption prediction

• Real-time energy prediction is difficult because providers mainly estimate host energy offline, workloads are nonlinear, and no single ML algorithm performs well across hosts.The passage notes that Google cluster nodes use no more than 60% CPU and 50% memory, motivating ensemble learning for accurate prediction.Source: §3.10 · pp. 27–28
• [84] combines four predictors to forecast VM energy efficiency in real time, but omits LLC and disk throughput and remains workload-specific.The ensemble uses moving average, exponential smoothing, linear regression, and double exponential smoothing, selecting predictions by mean absolute error across iterations.Source: §3.10 · p. 28

3.11. Prediction Accuracy in Auto-Scaling of web applications

• [85] uses a genetic algorithm to weight time-series models for auto-scaling, adapting to changing workloads and model types but incurring high time complexity that can violate SLAs.The approach can incorporate more advanced prediction models, but its response-time impact is a stated limitation.Source: §3.11 · pp. 28–29

3.12. Time-Series Prediction Data

• [86] dynamically adds or removes time-series predictors using scores, a threshold of 5, and a floor of 0, but its fixed parameters limit generalization beyond the chosen dataset.Dynamic adjustment is intended to respond quickly to nonlinear workload trend changes.Source: §3.12 · p. 29

3.13. Data Training

• Virtual-resource forecasting must dynamically follow changing workloads and select training data that strongly affects targets because nonlinear vCPU and vRAM demand makes prediction difficult.The section identifies dynamic prediction and impactful training-data selection as the two central forecasting requirements.Source: §3.13 · pp. 29–30
• [87] uses a Bayesian approach to select influential training variables for nonlinear virtual-resource prediction with fewer SLA violations, but excludes application combinations and high-level throughput and latency metrics.The method identifies dependencies systematically using data from Amazon, EC2, and Google, yet remains problem-specific and lacks generalization.Source: §3.13 · p. 30

3.14. VM Multi Resources

• [88] clusters VMs and assigns each cluster an Extreme Learning Machine predictor for multi-resource request forecasting, but fixing k-means at three clusters restricts flexibility.ELMs avoid several gradient-based training issues, and separate networks can be used for each cluster, but the cluster count is predetermined.Source: §3.14 · pp. 30–31

4.1. Performance and Online Profiling of Workload

• Future workload prediction should improve accuracy for nonlinear VM workloads while reducing time complexity through online profiling, resource-utilisation prediction, and semi-supervised classification with partial labels.Online profiling should consider CPU and memory utilisation, while semi-supervised learning may classify interactive and delay-insensitive VMs in large distributed data centres.Source: §4.1 · p. 31

4.2. Multiple Resource Usage in VM consolidation

• VM consolidation should detect overload using CPU, memory, and bandwidth jointly, while reducing training and migration delays through efficient multi-feature or single-feature DL models.LSTM and GRU models can use CPU-utilisation vectors to predict future states while avoiding some multiple-feature training overheads.Source: §4.2 · pp. 31–32

4.3. Cloud Network Traffic

• Resource allocation and thermal management should forecast disk throughput, CPU load, and host temperature to reduce migration delays, tail latency, cooling costs, and algorithm overhead.Single-feature GRU-based prediction and prior CPU estimation are proposed to support faster temperature reduction and prevent CPU overload.Source: §4.3 · p. 32, §4.4 · pp. 32–33

4.5. False Host Overloaded Detection

• VM migration should use dynamic thresholds and near-future degradation information so short-lived utilisation drops do not trigger unnecessary migration, while sustained degradation still prompts migration.This direction addresses unreliable migration caused by static overloaded-host thresholds.Source: §4.5 · p. 33

4.6. Energy metering at Software-Level

• Software-level energy management should cluster VMs into consumption states using correlated features, while power models should include memory, disk, and network and capture nonlinear relationships.Feature-selection methods and clustering ensembles could group VMs for joint management, potentially reducing response time and improving QoS.Source: §4.6 · pp. 33–34, §4.10 · p. 36

4.7. SLA-based VM Management

• SLA- and QoS-aware provisioning should combine dynamic consolidation, optimised ANN hyperparameters, reactive correction, peak-usage forecasting, ensemble learning, and request intervals matched to VM deployment time.These directions target nonlinear workloads, prediction error, response time, rejection rate, and SLA-sensitive resource demand.Source: §4.7 · p. 34, §4.8 · pp. 34–35

4.9. Varying Patterns of a Service Tenant in Resource Allocation

• Tenant and VM heterogeneity should be addressed through unsupervised or semi-supervised clustering, workload-specific model adaptation, feature selection, and hyperparameter optimisation using methods such as Grid Search and Bayesian Optimization.Clustering ensembles can improve VM grouping while evaluating clustering accuracy, time complexity, and resource usage.Source: §4.11 · pp. 36–37, §4.13 · p. 38, §4.14 · p. 38

4.12. Time-Series Prediction Data

• Future time-series research should develop generalised ensembles and hierarchical global-local models, and evaluate CNN, attention, dilated-convolution, and TCN architectures for cloud workload forecasting.These models are proposed to handle outliers, seasonality, local dependencies, and long-range temporal dependencies more effectively than traditional RNN approaches.Source: §4.12 · p. 37

Scope

• The review examines ML-based workload prediction, energy-consumption prediction, and related resource-management tasks using different ML methods.It covers the approaches used to address these challenges and their benefits and drawbacks.Source: §5 · pp. 38–39

Findings

• ML models can achieve diverse optimization goals and handle complex resource-management tasks in cloud computing.ML approaches also enable intelligent resource and application management.Source: §5 · p. 39

Future directions

• The paper identifies research gaps and proposes future directions to strengthen ML methods for resource management in cloud-based systems.Advanced reinforcement-learning and deep-learning techniques are highlighted as promising ways to improve system efficiency.Source: §5 · p. 39