A cloud-management framework reported lower average power consumption and higher simulated profit than OP-PMF− in comparative tests using two benchmark virtual-machine traces. Average power consumption with OP-PMF+ was almost 50% lower on Google Cluster Data (GCD) and up to 60.7% lower on PlanetLab (PL). Maximum average profit was reported as up to 49.72% for GCD and 51.18% for PL. The work was evaluated through experimental simulation, so the figures describe the selected model comparisons rather than a live cloud deployment.
A model built around prediction and pricing
OP-PMF was designed to use resource oversubscription and heterogeneous service pricing to maximize profitability and operational efficiency for industry cloud providers. The framework is organized into a Load Balancing Unit and a Resource Management Unit. It proactively predicts virtual-machine resource use with an adaptive ensemble, then groups predicted VMs by resource-use similarity using Fuzzy C-means clustering. The reported design uses that grouping alongside autoscaling.
Pricing is built into request handling. Requests are classified as DSM or BEM according to the selected pricing model: DSM allows higher payment for earlier completion, while BEM uses fixed charges based on resource requirements. That classification feeds request placement and the study’s constructed profit measure.
The test was synthetic by design
The evaluation compared OP-PMF with other methods in experimental simulation using GCD and PL. GCD contained 672,300 jobs on 12,500 servers over 29 days. PL supplied CPU-utilization traces for more than 11,000 VMs, measured every five minutes during 10 random days in March and April 2011.
The traces did not include user, deadline or request-charge information, so those inputs were generated for the simulation. The study set synthetic users at 60% of the total VM count and randomly mapped between one and five VMs to each user. DSM request share was varied across 20%, 40%, 60%, 80% and 100%, with a five-minute prediction and processing interval.
Strong results inside the model
On both traces, OP-PMF’s predicted CPU and memory use were reported to almost overlap actual use, while NN, RF, LR and SVM fluctuated over time. Overload prediction also produced high reported figures: correctly predicted overloads reached up to 99.6%, unpredicted overloads were at most 2.3%, and overload-prediction accuracy was approximately 99.94% on GCD and 99.91% on PL.
OP-PMF+ was reported to lower average total active-server percentage by up to 49% for GCD and 51% for PL against OP-PMF−. Against existing approaches, OP-PMF had a mean active-server value of 50%. For GCD, reported resource utilization was 60% higher than OP-PMF−, 37.9% higher than FF-WP, 21.25% higher than RF-WP and 18.75% higher than BF-WP.
The VM types and counts required were determined beforehand, and the paper described the resulting OP-PMF autoscaling as near-optimal. In the same comparison, average power consumption with OP-PMF+ was almost 50% lower for GCD and up to 60.7% lower for PL than with OP-PMF−.
The financial result rests on the simulation’s assumptions
Electricity cost was computed at $0.07 per kilowatt-hour, alongside a $0.25-per-hour DSM charge for early execution. At a 100% DSM share, average electricity cost with OP-PMF+ was reported as up to 53.3% lower for GCD and 55.56% lower for PL than with OP-PMF−.
The study estimated profit as electricity-cost reduction per hour using rDSM (%) × 0.25 × te. In the tests, rDSM ranged from 20% to 100%, and te, the earlier-execution time, was set to five minutes. OP-PMF+ always earned more in the reported comparisons. Its maximum average profit was reported as up to 49.72% for GCD and 51.18% for PL.
A result bounded by its test setup
The results have a narrow evidentiary boundary. They come from algorithmic simulations on GCD and PL traces, with synthetically generated user and request information, fixed pricing and timing assumptions, and specified comparison methods. The study reports no confidence intervals or inferential tests, so it offers no formal statistical estimate of uncertainty around the percentages. The findings therefore describe performance under the selected traces, server and VM settings, and DSM/BEM mixes; they do not by themselves establish the same power, cost or profit outcomes in a live cloud or under real customer charges and deadlines.
Paper data and sources
Original title: An Oversubscription and Service Pricing Exploitation-Based Profit Maximization Framework for Industry Cloud Resource Management
Authors: Deepika Saxena, Ashutosh Kumar Singh
Journal/Repository: IEEE Transactions on Services Computing 2024
Status: Peer-reviewed
First online: 2026-08-26
DOI: 10.1109/tsc.2024.3445379
Original paper · Full text