Source-linked AI summary
Machine Learning (ML)-Centric Resource Management in Cloud Computing: A Review and Future Directions
Tahseen Khan, Wenhong Tian, Rajkumar Buyya
TL;DR
Cloud resource management must address nonlinear workloads and limitations of static policies in dynamic environments. This paper reviews ML-based approaches across resource-management tasks, evaluates their benefits and drawbacks, and proposes future research directions. Its findings support ML for optimization and complex cloud-management tasks while identifying unresolved limitations in current methods.
Problem
Dynamic, nonlinear cloud workloads and static resource-management policies create challenges for efficient management across diverse tasks and environments.
Method
The paper reviews ML-based resource-management approaches, their challenges, benefits, limitations, experimental settings, datasets, and performance improvements.
Results
The findings show that ML models can support optimization goals and complex tasks in cloud-computing systems, opening an avenue for intelligent resource and application management.
Takeaways & Limitations
The paper identifies advanced techniques such as reinforcement learning and deep learning as promising directions for intelligent resource management.
Takeaways & Limitations
Reviewed approaches retain limitations including restricted metrics, fixed ensemble parameters, and challenges applying a single ML model to nonlinear workloads.
Abstract
from arXiv · showhide
Cloud computing has rapidly emerged as model for delivering Internet-based utility computing services. In cloud computing, Infrastructure as a Service (IaaS) is one of the most important and rapidly growing fields. Cloud providers provide users/machines resources such as virtual machines, raw (block) storage, firewalls, load balancers, and network devices in this service model. One of the most important aspects of cloud computing for IaaS is resource management. Scalability, quality of service, optimum utility, reduced overheads, increased throughput, reduced latency, specialised environment, cost effectiveness, and a streamlined interface are some of the advantages of resource management for IaaS in cloud computing. Traditionally, resource management has been done through static policies, which impose certain limitations in various dynamic scenarios, prompting cloud service providers to adopt data-driven, machine-learning-based approaches. Machine learning is being used to handle a variety of resource management tasks, including workload estimation, task scheduling, VM consolidation, resource optimization, and energy optimization, among others. This paper provides a detailed review of challenges in ML-based resource management in current research, as well as current approaches to resolve these challenges, as well as their advantages and limitations. Finally, we propose potential future research directions based on identified challenges and limitations in current research.
1. Introduction
Cloud resource management must handle variable workloads, balancing utilization, QoS, energy use, and costs. This review surveys ML-based approaches, their challenges, benefits, limitations, and future directions.
- Variable VM workloads can over- or underutilize servers, causing inconsistent QoS, unbalanced energy use, and SLA violations.
- Optimal VM-to-server mapping is an NP-complete problem that must balance resource use with QoS requirements and data-center benefit.
- The review examines ML-based provisioning, VM consolidation, thermal prediction, and other resource-management tasks, including experimental settings and datasets.
- Static threshold policies can create unnecessary migrations, overhead, and response delays when workloads change rapidly.
- The paper identifies advantages and drawbacks of current methods and proposes future research directions based on unresolved challenges.
2.1. Cloud Computing
Cloud computing provisions computing resources over the Internet and supports centralized application deployment, computation offloading, flexible provisioning, and cost-effective delivery.
- Cloud computing provides resources such as memory, CPU, bandwidth, disk, and applications or services over the Internet.
- Its client-server architecture supports centralized application deployment and computation offloading.
- Cloud computing is described as cost-effective and flexible in delivering, provisioning, and detaching services from related technologies.
2.2. Core features of cloud computing •
Cloud computing combines on-demand access, broad reach, pooled resources, elasticity, measured services, multitenancy, and auditability.
- On-demand self-service lets clients request services as needed and pay through an online control center.
- Broad network access enables resources and services to be accessed from multiple locations through standard mechanisms.
- Resource pooling presents resources as a blended pool whose physical service locations are hidden from clients.
- Rapid elasticity allows resources and services to scale up or down according to demand.
- Measured service supports automatic control, monitoring, optimization, and documentation, while multitenancy requires policy-driven separation and governance.
- Auditability and certifiability require logs and trails for assessing compliance with laws and policies.
2.3. Cloud computing service models •
Cloud service models differ in how much application, platform, and infrastructure responsibility remains with the user.
- Software as a Service provides access to provider-hosted applications through web portals.
- Platform as a Service supplies networks, servers, and operating systems for clients to build and configure applications.
- Infrastructure as a Service provides basic infrastructure such as processors, networks, and storage for user-created applications.
2.4. Deployment models for cloud computing •
Cloud computing deployment models include public, private, community, and hybrid clouds, distinguished by ownership, sharing, security, cost, and interoperability.
- Public clouds provide Internet-based services under predefined rules and offer resource choices while maintaining QoS.
- Private clouds provide organizational benefits with fewer security concerns but impose high establishment and management costs.
- Community clouds let organizations share infrastructure under common criteria, providing cost sharing and protection.
- Hybrid clouds combine independent public, private, or community clouds while requiring interoperability and portability across constituent services.
2.5. Machine Learning
Machine learning trains models from data to predict outputs or recognize patterns, with supervised, unsupervised, semi-supervised, and reinforcement learning addressing different information settings.
- Machine learning constructs predictive algorithms from training data and statistical techniques rather than explicitly programming every input-output case.
- Supervised learning maps input features to labeled outputs for prediction.
- Unsupervised learning learns data distributions without labels, including clustering VMs by resource-use patterns.
- Semi-supervised learning combines labeled and unlabeled data to improve classification, regression, clustering, or dimensionality reduction.
- Reinforcement learning trains agents through environmental interaction, balancing exploration and exploitation using rewards and sanctions.
2.6. Optimization objective in machine learning
Machine-learning methods formulate objective functions and use numerical optimization to learn mappings, clusters, projections, policies, or semi-supervised classifiers.
- Machine-learning algorithms optimize objective functions using numerical optimization methods.
- Supervised learning: Supervised learning minimizes training loss to find an optimal mapping from feature vectors to labels.
- Supervised learning: Regularization, including the l2-norm, is added to objective functions to prevent overfitting, with γ calculated by cross-validation.
- Unsupervised learning: K-means minimizes within-cluster variance by assigning samples to cluster centers.
- Unsupervised learning: PCA projects data into lower-dimensional coordinates while minimizing reconstruction error and retaining original information as much as possible.
- Reinforcement learning: Reinforcement learning maximizes cumulative return through a policy mapping states to actions.
- Semi-supervised learning: Semi-supervised S3VM optimization uses labeled and unlabeled samples with slack variables and penalty constraints for classification.
3. Challenges, state-of-art research and their limitations
The review identifies workload, profiling, and multi-resource challenges in ML-based cloud resource management, surveys existing approaches, and proposes future research directions.
- The review examines challenges, current solutions, advantages, and limitations in ML-based resource management.
- Performance and online profiling of workload: Offline workload profiling is often infeasible because workloads may become available only after VMs enter production, while online profiling is difficult.
- Performance and online profiling of workload: Microsoft Azure’s prediction system learns from historical data and provides online forecasts to multiple resource managers through a REST API.
- Multiple Resource Usage in VM Consolidation: VM consolidation reduces energy use by packing VMs onto fewer hosts, but current-utilization decisions can cause unnecessary migrations, power transitions, or future overutilization.
- Multiple Resource Usage in VM Consolidation: Linear Regression predicted future VM resource utilization from historical Plane workload traces to support dynamic host selection for consolidation.
- VM Placement: VM placement remains challenging because workloads vary over time and may require jointly considering CPU demand and network bandwidth.
- VM Placement: A network-aware predictive placement heuristic used predicted CPU utilization and network bandwidth to reduce energy consumption and SLA violations.
3.4. Host Temperature
Machine-learning-based resource management addresses thermal, energy, utilization, SLA, and provisioning challenges caused by variable cloud workloads. Reviewed approaches use prediction to guide scheduling and allocation, but retain limitations involving detection criteria, measurement, forecasting, and tenant classification.
- Host Temperature: 6.5°C peak-temperature and 34% energy-consumption reductions were reported for thermal-aware predictive scheduling compared with existing algorithms.The approach predicted host temperature and migrated VMs to fewer hosts to reduce energy consumption.
- False Host Overloaded Detection: False overloaded-host detections arise when utilization briefly exceeds a threshold before declining rapidly, complicating VM-migration decisions.Longer-lasting load degradation may instead require VM migration.
- False Host Overloaded Detection: VM consolidation used current and predicted utilization to identify hosts overloaded now and hosts predicted to become overloaded later.The reviewed strategy did not address hosts overloaded currently but predicted not to remain overloaded.
- Energy Metering at Software-Level: Software-level VM energy measurement remains difficult because server meters cannot isolate the energy consumed by individual VMs.This complicates energy-aware VM consolidation and energy prediction.
- SLA-based VM Management: Historical CPU utilization and SLA penalties supported ANN-based forecasts for complex placement decisions across compute-intensive and transactional applications.The approach predicted VM CPU usage for the next two hours, but forecasts sometimes diverged from actual utilization.
- QoS-Aware Resource Provisioning: Static resource allocation wastes resources during low demand, while ARIMA-based provisioning estimates future requests and accounts for response time and rejection rate.The ARIMA approach used a fixed forecasting interval, which could leave extra time affecting response time when VM deployment is faster.
3.10. Single ML model in energy consumption prediction
Energy-use prediction in cloud environments is difficult because workload behavior is nonlinear and real-time estimation remains challenging. The reviewed approaches use ensembles or workload-specific models, but their generalization is limited by omitted metrics, fixed parameters, and narrow dependencies.
- Real-time energy estimation remains challenging for cloud providers during application execution.
- Ensemble learning combines multiple predictors to improve energy-efficiency forecasting for VM resources in real time.One reviewed method uses moving average, exponential smoothing, linear regression, and double exponential smoothing, selecting predictions using mean absolute error.
- The reviewed ensemble method omits LLC and disk-throughput metrics, although these affect host energy consumption at the VM level.
- Prediction accuracy is workload-specific, so the selected model does not generalize across interactive and batch workloads.
- Dynamic ensemble updating can respond to nonlinear workload trends, but fixed threshold and floor parameters produced a dataset-specific, non-generalized approach.
4. Future Research Directions
Future research directions focus on making ML-based resource management more predictive, adaptive, and comprehensive across workloads, resources, and energy use. Key priorities include dynamic provisioning and consolidation, richer resource models, improved time-series methods, and ensemble-based approaches.
- Dynamic resource management: Dynamic VM consolidation should combine CPU, memory, and bandwidth utilization with current and future resource estimates to detect overloaded hosts more reliably.Existing approaches may rely on CPU thresholds, while multi-resource detection is proposed as a broader direction.
- Dynamic resource management: Dynamic resource-utilization thresholds should use near-future data to avoid unnecessary VM migrations caused by short-lived utilization changes.Fixed overloaded-host thresholds can produce unreliable migration decisions.
- Energy management: VM-level energy management may use feature selection and clustering to classify VMs by energy state without requiring host-level features.The proposed categories include low, moderate, and critical energy consumption, potentially enabling grouped management and improved QoS.
- QoS-aware provisioning: Future provisioning research should address reactive demand, prediction error, dynamic estimation intervals, and more sophisticated or ensemble ML models for nonlinear workloads.Suggested models include XGBoost, LSTM, and GRU, while dynamic intervals can adapt to VM deployment time.
- Workload prediction: Time-series research should develop generalized ensembles, hierarchical global-local models, and architectures that capture both local dependencies and long-term seasonality.The review identifies limitations of global neural networks and traditional RNNs for outliers and seasonal patterns.
- Multi-resource demand: Clustering ensembles could improve multi-resource VM-demand prediction by combining multiple clustering algorithms into a consensus solution.This is proposed as an alternative to k-means, which may constrain cluster counts and misplace VMs.
5. SUMMARY AND CONCLUSIONS
The paper reviews how machine learning is being applied to workload, energy, and related cloud resource-management tasks, together with their benefits and drawbacks. It concludes that ML can support optimization and complex management tasks, while advanced methods such as reinforcement learning and deep learning remain promising directions.
- Summary: The review examines recent ML-based approaches for cloud resource-management challenges, including their benefits, drawbacks, and proposed future directions.It is intended to help cloud researchers understand resource management and the role of ML techniques.
- Conclusions: ML models can address complex cloud-computing tasks and achieve various optimization goals across resource-management applications.The paper also identifies progress in current ML approaches and research gaps in the field.
- Future directions: Advanced techniques such as reinforcement learning and deep learning are identified as promising approaches for intelligent resource management.