An Optimum Compute Resources Consolidation Framework for Cloud Data Center

Abstract

Cloud computing offers an efficient alternative for business enterprises as compared to traditional models for computing and data storage needs. The dynamic workload on cloud servers is one of the major reasons for the ineffective utilization of cloud resources. Therefore, a larger number of servers are active in cloud data centers (CDCs) to satisfy the cloud users’ demands, which severely enhances energy consumption and heat dissipation. Hence, effective resource management within the CDCs has become crucial for cloud service providers. To meet users’ requests of cloud services, it is essential to optimally allocate requested resources on physical machines (PMs) through virtual machines (VMs). The dynamic nature of workloads complicates initial VM allocation, which often results in the overutilization or underutilization of PMs. This leads to performance degradation, wastage of resources, increased operational costs, higher active servers, and energy consumption. Therefore, to address these challenges, effective resource management strategies are required to ensure the effective utilization of PM resources. With an intent to achieve e!ective resource management, this thesis proposed solutions to accurately predict the resource usage of machines and e!ectively utilize resources to balance the load of PMs, as well as reduce the total energy consumption of a data center. Firstly, in order to e!ectively utilize the resources of PMs in a data center, a hybridizing approach leveraging Gaussian Mixture Model (GMM) and Long Short-Term Memory (LSTM) model is proposed to predict resource usage of heterogeneous PMs. This research aims to capture the heterogeneity of available PMs in a data center using GMM based on mean CPU usage and memory usage of machines while capturing temporal dependencies and patterns in resource usage data using optimal hyperparameters of the LSTM model that enable more accurate prediction. Next, to capture both long-range and short-term dependencies, as well as input se quences of data to predict the resource usage of PMs, rigorous experiments are carried out. Attention-based mechanism models, Transformer and Informer, along with LSTM, are employed for this purpose. Moreover, to rationally select heterogeneous machines based on mean CPU, memory, and hard disk usage, the Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH) algorithm is utilized. To e!ectively utilize the resources of PMs in the data center, it is essential that requested VMs are allocated to PMs in a manner that maximizes resource usage and reduces the overall energy consumption of the data center. To achieve this, we employed Best Fit, First Fit, Best Fit Decrease, First Fit Decrease, and Genetic Algorithm for optimal VM allocation (VMA) to minimize active PMs as an objective function and identify their effect on energy consumption. Lastly, a three-tier architecture is proposed for the allocation and dynamic consolidation of VMs using current as well as predicted resource usage of machines in a CDC. The first tier carries out a Non-Dominated Sorting Genetic Algorithm (NSGA-II) for VMA by considering their resource constraints. The second tier employs LSTM for resource usage prediction of VMs, which enables the computation of PM current as well as predicted resource usage. The third tier aims to achieve dynamic consolidation of the VMs by detecting overloaded and underloaded PM using threshold-based criteria, VM selection using a proposed novel strategy based on Pareto-front and VM placement using NSGA-II. The objective of this architecture is to provide a holistic approach that reduces the active number of PMs, the number of migrations, and the energy consumption of the CDC with aminimumpredictionerrorrate. Google Cluster Trace usage dataset (GCT) is utilized to assess and validate the techniques employed in our study. A Sum-Average (SA) algorithm is proposed for data preprocessing to extract PM resource usage from task resource usage. It prepares the dataset and makes it suitable for providing input into clustering and prediction models.