PaperlayerView exact revision on arXiv

Source-linked AI summary

An Oversubscription and Service Pricing Exploitation-Based Profit Maximization Framework for Industry Cloud Resource Management

Deepika Saxena, Ashutosh Kumar Singh

arXiv:2608.25712v1cs.DC

TL;DR

Industry cloud providers face resource wastage, excess power consumption, and high electricity costs from over-reserved resources and underutilized servers. OP-PMF combines proactive VM prediction, fuzzy clustering, oversubscription, and heterogeneous pricing models; simulations report improved resource utilization and reduced operational costs.

  • Problem

    Industry cloud users over-reserve CPU and memory because actual resource demand is unknown, leaving servers underutilized and increasing power consumption and operational costs.

  • Method

    OP-PMF proactively predicts VM resource utilization with an adaptive ensemble model, clusters predicted VMs using Fuzzy C-means, and exploits oversubscription with Delay Sensitive and Best-Effort pricing models.

  • Results

    OP-PMF improves resource utilization and reduces power-related operational costs in simulations against alternative resource-allocation approaches.

  • Takeaways & Limitations

    Oversubscription combined with proactive scaling and pricing models supports profit maximization while consolidating workloads on fewer active servers.

Abstract

from arXiv · show

This article proposed a novel industry cloud resource management framework that exploits resource oversubscription and heterogeneous service pricing models to maximize profitability and operational efficiency for industry cloud providers. The framework proposes an adaptive ensemble machine learning driven prediction model for proactive estimation of resource utilization of Virtual Machines (VM)s based on previous resource utilization of respective users' VMs to minimize resource wastage due to oversubscription by them. Accordingly, the VMs having similar predicted resource usage are grouped using Fuzzy C means clustering. This helps to determine the required number of VMs with specific configuration to be deployed before executing user requests. Concurrently, the framework incorporates two distinct categories of cloud service pricing models, namely the Delay Sensitive Model and the Best-Effort Model. Accordingly, the user requests are classified and executed by selecting the most suitable VMs, with the goal of maximizing revenue and reducing electricity costs in cloud data centers (CDCs). Experimental simulation and comparison against state-of-the-art methods, using two benchmark VM traces, validates the performance of proposed framework. It significantly reduces electricity bills by 55.56 percentage, power consumption and active servers by up to 60.7 percentage and 51 percentage, respectively, while improving resource utilization and profits by up to 60 percentage and 51.18 percentage, respectively

I. INTRODUCTION

Industry cloud providers face high electricity costs and low server utilization because users over-reserve resources under uncertain demands. OP-PMF addresses this through oversubscription, adaptive VM prediction, clustering, and heterogeneous pricing models.

  • Motivation: Server utilization often remains between 10% and 50%, causing resource wastage and excess power consumption from idle or under-utilized servers.Users overestimate CPU and memory needs because actual demand is unknown before execution and may vary over time.
  • Motivation: Oversubscription allocates VM resources beyond physical server capacity by relying on non-simultaneous full utilization across VMs.The approach uses statistical multiplexing to improve utilization, reduce power consumption, and support scalability.
  • Motivation: 69.4% reduction in power consumption and operational cost is reported for optimal oversubscription scaling against resource allocation without oversubscription.The reported reduction is associated with consolidating workloads on fewer active servers.
  • Proposed direction: OP-PMF proactively predicts VM resource usage with adaptive ensembles, clusters VMs by predicted usage using Fuzzy C-means, and determines VM quantities and configurations for upcoming workloads.The framework also exploits heterogeneous pricing models to coordinate resource management with profit maximization.
  • Evaluation: The proposed framework is evaluated on two real benchmark datasets and compared with state-of-the-art approaches across prediction, utilization, server, power, and electricity-bill metrics.The stated evaluation covers both resource-management performance and cost-oriented outcomes.
  • Research gap: Existing approaches primarily emphasize load balancing and task scheduling for energy efficiency while overlooking joint use of oversubscription and heterogeneous pricing for profit maximization.OP-PMF is introduced to integrate these aspects while optimizing costs and profits within the stated cloud-management scope.

III. PROBLEM FORMULATION

The formulation seeks to distribute physical resources among users’ VMs to maximize provider profit while executing requests within resource, placement, timing, and service constraints. It minimizes electricity and server-related costs while improving resource utilization.

  • Objective: The problem distributes servers’ physical resources among users’ VMs to maximize provider profit while serving requests without violating SLA terms.Requests specify demanded CPU and memory capacity together with execution start and end times.
  • Constraints: Each VM must be deployed on one server, while request and aggregate resource capacities must remain within available VM and server capacities.These constraints govern individual placement, per-request feasibility, and total resource distribution.
  • Constraints: Each request’s execution time must not exceed its deadline, and its CPU and memory requirements must not exceed available capacity.The formulation therefore combines temporal feasibility with resource-capacity constraints.
  • Objective: The objective minimizes electricity expenditure and active-server counts while maximizing data-center resource utilization.Resource utilization is improved by reducing physical capacity allocated beyond actual request requirements.
  • Pricing: Service pricing model and request price are represented by PM and rPM, respectively, within the time-slot resource-management formulation.The formulation explicitly connects resource distribution with service-pricing terms.

IV. PROPOSED FRAMEWORK

OP-PMF combines request distribution and backend resource management to prepare VM capacity for future demand. It predicts VM usage, clusters predicted VMs by similarity, and uses the resulting groups for proactive autoscaling and request execution.

  • Framework architecture: OP-PMF comprises a Load Balancing Unit and a Resource Management Unit for handling user requests and backend resource-scaling knowledge.The units operate across users’ VMs, hosted servers, available resources, expected load, and next-slot deployment needs.
  • Framework architecture: The load-balancing process distributes user requests among VMs hosted on servers and classifies requests according to their selected pricing model.Pricing-based classification occurs before request execution.
  • Resource prediction: An Adaptive Ensemble Model combines predicted outputs from multiple base learners to estimate future resource usage for different VMs.Periodic training and retraining update the base learners and support prediction of VM resource usage.
  • VM autoscaling: Predicted VMs are grouped into K clusters according to their predicted resource usage for autoscaling.The clusters represent VM groups with similar estimated usage and support subsequent deployment decisions.
  • VM autoscaling: The resulting VM clusters help estimate the exact number and types of VMs needed to serve user demand in the next time-slot.This connects prediction and clustering to proactive resource preparation.

A. Adaptive Resource Prediction Unit

The resource prediction unit prepares VM utilization histories, generates adaptive weighted ensembles from multiple base learners, and periodically updates predictions using observed performance.

  • The ensemble is periodically recreated through self-adaptive weight allocation during training or retraining of the real-time resource predictor.
  • The unit aggregates VM resource-usage samples, normalizes them to [0, 1], and forms a learning window for future prediction.
  • Base learners including Linear Regression, Support Vector Machine, Neural Network, and Random Forest generate VM resource predictions and RMSE errors.
  • Prediction errors are normalized, then converted into learner-specific weights so better-performing learners contribute more strongly to each VM’s ensemble.
  • Algorithm 1 applies VM prediction across active VMs, with complexity described in terms of the number of VMs and base learners.

B. Leveraging Cloud Oversubscription

The framework uses predicted VM resource requirements to group similar workloads with Fuzzy C-means and select VM configurations and quantities for the next time-slot.

  • Predicted VMs are clustered by estimated resource usage to determine the VM sizes and numbers needed for the subsequent time-slot.
  • Fuzzy C-means assigns membership values according to distances between predicted VMs and cluster centers, with smaller distances indicating stronger membership.
  • The clustering objective makes resource usage within clusters similar while keeping different clusters separated.
  • VM configuration is selected by comparing each cluster’s predicted resource range with available small, medium, large, and extra-large VM capacities.
  • For a selected configuration, the required VM count equals the total number of predicted VMs in the corresponding cluster.

C. Exploiting Service Pricing Models

Requests are divided by pricing model, then assigned to faster or slower servers according to deadline sensitivity, processing requirements, and execution cost.

  • The framework categorizes requests into Delay Sensitive Model and Best-Effort Model classes based on the user-selected service pricing model.
  • Delay Sensitive Model: DSM requests require deadline-bound execution, and users agree to pay more for completion before the deadline.
  • VM placement is permitted only when the VM’s CPU and memory requirements do not exceed the available capacity of the target server.
  • Delay Sensitive Model: DSM requests are ordered by increasing deadline and assigned to high-processing-speed servers satisfying the resource constraints.
  • Best-Effort Model: BEM workloads are assigned preferentially to lower-speed active servers, where lower execution cost is used to reduce power consumption and operational cost.

V. OPERATIONAL DESIGN AND ILLUSTRATION

OP-PMF predicts and clusters VM demand, estimates upcoming configurations, categorizes requests, and allocates them to servers under pricing and capacity constraints.

  • Operational initialization: The operational procedure initializes user, server, and VM sets, then repeats resource-management steps across time-slots.
  • Prediction and scaling: Historical CPU and memory utilization enters the VM prediction unit, and predicted VMs are clustered to estimate next-slot VM requirements.
  • Request preparation: Requests are separated into DSM and BEM lists before available servers are sorted for allocation.
  • Pricing-aware allocation: DSM requests are assigned to suitable high-speed servers, while BEM requests are assigned to suitable low-speed servers.
  • Complexity: The complete OP-PMF procedure has time-complexity O(P 2Q2x2ZC).The stated complexity aggregates prediction, clustering, scaling, sorting, and VM-allocation operations.

B. Numerical Illustration

The numerical illustration applies OP-PMF to seven VMs across Best-Effort and Delay Sensitive pricing models, comparing its consolidated placement with classical VM placement. It reports lower power and electricity costs alongside an earned-profit estimate.

  • Scenario and prediction: Seven VMs from two pricing-model scenarios are assigned predicted resource usage and grouped into a cluster for placement.The scenario includes Best-Effort and Delay Sensitive requests, with VM configurations referenced from the paper’s configuration tables.
  • VM placement: 35%-50% average server utilization motivates consolidating the seven VMs on one S1 server.The illustration places the clustered VMs on one server with two processing elements.
  • OP-PMF outcome: 2158.7 KW power consumption and 151.109 $/H electricity expenditure are reported for OP-PMF at 50% resource utilization.These values are obtained after applying the cited power and utilization equations.
  • Cost comparison: 49.4% electricity-bill reduction is reported for OP-PMF compared with classical placement.Classical placement costs 302.208$/H on two S1 servers or 247.746 $/H on one S2 server.
  • Profit outcome: 42.2% of VMs receive earlier-execution privilege under the pricing mix, producing an earned profit of 1.76$ per hour.Industry A uses Best-Effort service and industry B uses Delay Sensitive service, with earlier execution assumed to be 10 minutes.

VI. PERFORMANCE EVALUATION AND COMPARISON

The evaluation uses Google Cluster and PlanetLab VM traces to assess prediction, resource management, and cost-related performance against existing approaches. OP-PMF combines online prediction and scaling within a simulated data-center environment.

  • Experimental platform: The simulation uses three server types and four VM configurations based on real IBM and Dell server specifications.Experiments run in Python 3 on a machine with two Intel Xeon Silver 4114 CPUs, 40 cores, and 128 GB memory.
  • Datasets: Google Cluster and PlanetLab traces provide the two real workloads used to evaluate VM resource utilization.Google Cluster includes 672,300 jobs on 12,500 servers over 29 days; PlanetLab contains CPU utilization traces for more than 11K VMs.
  • Experimental assumptions: The original traces lack users, deadlines, and execution charges, so the evaluation synthetically assigns users, VM ownership, and request pricing classes.Users are generated as 60% of the total VMs, with each user owning between one and five VMs.
  • Baselines: OP-PMF is compared with SBA, THR-P, LR-P, OP-MLB, First-Fit, Random-Fit, and Best-Fit approaches across multiple performance metrics.The comparison includes prediction, resource utilization, active servers, power saving, and electricity-bill optimization.
  • VM prediction: Predicted CPU and memory usage nearly overlap actual usage for both traces, while neural network, Random Forest, Linear Regression, and Support Vector Machine predictors fluctuate over time.The adaptive ensemble periodically retrains using live and historical usage patterns and correlations.

2) Number of Overloads:

This section evaluates overload prediction, active-server reduction, and resource utilization across Google Cluster and PlanetLab workloads. OP-PMF reports high overload-prediction accuracy and lower active-server percentages than its comparison settings.

  • Overload prediction: Up to 99.6% of overloads are correctly predicted, while unpredicted overloads remain at or below 2.3% across VM traces.Unpredicted overloads are reported as independent of the ratios of Delay Sensitive to Best-Effort requests.
  • Number of Active Servers: 49% and 51% reductions in average active-server percentage are reported for Google Cluster and PlanetLab, respectively, with OP-PMF enabled.The comparison varies the percentage of Delay Sensitive requests and contrasts OP-PMF+ with OP-PMF−.
  • Number of Active Servers: OP-PMF shows a mean active-server value of 50% for both workloads compared with First-Fit, Random-Fit, Best-Fit, THR-P, and LR-P.The reported stability is based on the comparison in Fig. 11.
  • Resource Utilization: OP-PMF improves average resource utilization by 60%, 37.9%, 21.25%, and 18.75% over OP-PMF−, First-Fit with prediction, Random-Fit with prediction, and Best-Fit with prediction for Google Cluster traces.The passage reports the same comparison structure for PlanetLab traces but truncates its final values.
  • Resource Utilization: For Google Cluster traces, OP-PMF improves median-quartile resource utilization by up to 81% over SBA, OP-MLB, Best-Fit, and First-Fit.The boxplot represents resource-utilization distributions through quartiles and whisker ranges.

3) VM Scaling:

OP-PMF predicts future VM resource utilization to scale the required VM types before request execution, then evaluates power and electricity-cost effects against alternatives. The framework reduces power and electricity costs while increasing profit across the benchmark traces.

  • VM Scaling: OP-PMF predicts future VM utilization and determines the required number of VSM, VME, VLA, and VXL instances before requests execute.The approach compares estimated and actual VM counts during execution.
  • Power Consumption: 60.7% lower power consumption is achieved by OP-PMF+ than OP-PMF− for Planet Lab VM traces.For Google Cluster traces, OP-PMF+ uses almost 50% less power than OP-PMF−.
  • Electricity Cost: 55.56% lower electricity cost is achieved by OP-PMF+ than OP-PMF− for Planet Lab traces at 100% rDSM.The corresponding reduction for Google Cluster traces is up to 53.3%.
  • Earned Profit: 51.18% maximum average profit is achieved for Planet Lab traces, compared with 49.72% for Google Cluster traces.The profit calculation uses electricity-cost reduction and earlier execution of delay-sensitive requests.

D. Trade-off between Earned Profits and Power Consumption

The framework examines the balance between profit maximization and power minimization across Google Cluster and Planet Lab traces. Its evaluation reports improvements in electricity bills, power consumption, active servers, resource utilization, and earned profits, while identifying VM failover as future work.

  • Trade-off between Earned Profits and Power Consumption: OP-PMF evaluates profit and power-consumption trade-offs for both Google Cluster and Planet Lab VM traces.The reported balance links resource-use and power optimization with increased execution of requests under DSM pricing.
  • Results: OP-PMF reports improvements in electricity-bill reduction, power consumption, active servers, resource utilization, and earned profits.The framework is evaluated through simulations and comparisons with state-of-the-art approaches using benchmark VM traces.
  • Future Work: VM failovers remain future work intended to improve continuity and resilience when VM failures occur.The paper proposes incorporating robust failover mechanisms for seamless operation and improved user experience and performance.
Loading 2608.25712v1…