Resource Allocation in The Cloud Environment with Supervised Machine learning for Effective Data Transmission

Mohammad Ebadi, Yu Wang, Khoshal Rahman Rahmani, Musawer HakimiJournal of Computer Science and Technology Studies 2024 · DOI

TL;DR

The paper introduces BRoML (Backtracking Regularized Optimized Machine Learning), a framework designed to solve the "Goldilocks" problem of cloud resource allocation: ensuring there is neither too much nor too little computing power for 5G applications. By combining historical backtracking with regularization and genetic optimization, the system achieves higher resource utilization and lower latency compared to traditional models like Support Vector Machines (SVM) and Random Forest (RF).

What Problem Does This Paper Address?

In the era of 5G, cloud environments face highly volatile demands. Traditional resource allocation—assigning virtual machines, storage, and bandwidth—often relies on static models that fail to adapt to real-time traffic spikes or drops. This leads to two major inefficiencies:

  1. Overprovisioning: Wasting expensive resources and energy.
  2. Underutilization: Causing high latency and poor user experience.

The authors aim to create a "responsive" infrastructure that can predict future demand ($P(t)$) and adjust available resources ($R(t)$) dynamically, ensuring data transmission remains seamless even as workloads fluctuate.

Methods

The core contribution is the BRoML framework, which integrates three distinct mathematical and algorithmic strategies:

  • Backtracking Mechanism: Instead of just looking at the current moment, the system uses a backtracking coefficient ($\alpha$) to blend historical demand patterns ($D_{history}$) with current demand ($D(t)$). This helps the system "remember" recurring patterns.
  • Regularization: To prevent the machine learning model from "overfitting" (reacting too wildly to noise or one-off anomalies), the authors incorporate L2 regularization. This adds a penalty term to the prediction, ensuring the model remains generalized and stable.
  • Genetic Optimization: The framework uses a genetic algorithm—complete with populations, crossover, and mutation—to iteratively "evolve" the best resource allocation configurations. This search process finds near-optimal solutions that minimize the gap between predicted and actual demand.
  • Supervised Learning: The optimized values are used to train a predictive model (linear regression base) that classifies and forecasts resource needs.

Datasets

The study utilized a publicly available 5G Quality of Service (QoS) dataset sourced from Kaggle.

  • Size: 400 samples.
  • Features: The dataset includes variables such as Timestamp, User ID, Application Type, Signal Strength, Latency, Required Bandwidth, and Allocated Bandwidth.
  • Preprocessing: Data was normalized to ensure features were on a comparable scale before being fed into the BRoML pipeline.

Key Insights and Ideas

  • The Latency "Sweet Spot": The authors propose a Latency Optimization Theorem stating that optimal latency is achieved only when the predicted future demand ($P(t)$) perfectly matches the current demand ($D(t)$). The goal of the ML model is to minimize the absolute difference between these two.
  • Responsiveness as a Derivative: The paper defines system responsiveness as the rate of change of available resources relative to demand ($dt/dR$). This mathematical framing treats the cloud as a dynamic fluid system rather than a static bucket of parts.
  • Evolutionary Efficiency: By using genetic optimization, the system doesn't just "guess" an allocation; it simulates thousands of possibilities and selects the "fittest" one, mimicking natural selection to handle the complexity of 5G networks.

Key Results

In head-to-head simulations against traditional machine learning schemes, BRoML consistently came out on top:

  • Resource Utilization: BRoML achieved 88.2% utilization, outperforming SVM (85.5%) and RF (87.8%).
  • Latency: BRoML reduced delay to 19.4 ms, compared to 22.1 ms for SVM and 20.5 ms for RF.
  • Throughput: The framework supported a data transfer rate of 160.8 Mbps, significantly higher than SVM’s 155.3 Mbps.
  • Overall Efficiency: The system maintained an average efficiency score of 0.93 (on a 0-1 scale) across various experimental runs.

Limitations and Future Work

While the results are promising, there are a few areas for further exploration:

  • Dataset Scale: The study relied on a relatively small dataset of 400 samples. Testing on massive, "wild" datasets from tier-1 cloud providers would be necessary to prove industrial-scale robustness.
  • Computational Overhead: Genetic algorithms can be computationally expensive. The paper does not deeply discuss the "cost of the cure"—i.e., whether the energy saved by better resource allocation is offset by the energy required to run the optimization itself.
  • Real-world Implementation: The current findings are based on simulation analysis. The authors note that future work should involve real-world deployments to validate the framework under unpredictable physical network conditions.

Summary updated

Abstract

Resource allocation in the cloud environment for 5G applications can be explained by referring to the strategic distribution and necessary assignment of computing resources such as virtual machines, storage, and network bandwidth that meet the dynamic demands of applications and services. The framework proposed is on resource allocation in the cloud environment by BRoML for 5G applications. In the proposed BRoML model, the Backtracking Regularized model is incorporated for the effective calculation of the resources in the cloud environment. The optimization is performed for the effective computation of resources in the cloud environment through the computed resources. Using the estimated optimized values, a machine learning model can be trained and tested to classify resource allocation. In this regard, the simulation analysis is compared to BRoML with traditional schemes like SVM and RF. The result shows that BRoML has a higher resource utilization while exhibiting lower latency, higher increased throughput, and a better efficiency score overall. Machine learning techniques and optimization mechanisms give flexibility and intelligence to BRoMl in solving resource allocation issues within cloud computing. These results reinforce the view that BRoML can create a strong impact on the development process of cloud computing with its dynamic, intelligent solution in resource allocation optimization under various scenarios.
BibTeX
@article{ebadi2024resource,
  author = {Mohammad Ebadi and Yu Wang and Khoshal Rahman Rahmani and Musawer Hakimi},
  title = {{Resource Allocation in The Cloud Environment with Supervised Machine learning for Effective Data Transmission}},
  journal = {Journal of Computer Science and Technology Studies},
  year = {2024},
  doi = {10.32996/jcsts.2024.6.3.3},
}
Figure