A COMPARATIVE STUDY OF ADVANCED LOAD BALANCING ALGORITHMSIN CLOUD COMPUTING ENVIRONMENTS

Abstract

Round Robin (RR) and First-Come First-Served (FCFS) scheduling algorithms have been designed with the assumption that workloads in cloud systems are uniform, which is not the case with today's cloud infrastructure, as it exposes many weaknesses of these algorithms. In this research, testing will be conducted on 10 different load balancing algorithms from 4 different groups: Artificial Intelligence(AI) and Deep Learning(DL); Nature-Inspired Metaheuristic Algorithms (NIMA); Game Theory Based Load Balancers (GT); and Traditional Load Balancers (LB). For this study, Google Cluster Trace data (from the Google data center) will be used to validate the performance of the aforementioned algorithms. BiLSTM-Attention reached 94.3% classification accuracy and 0.97 Area Under The Curve (AUC); SLADRO obtained 92% CPU Utilization and decreased Idle Power Consumption by 27.5%; these numbers are very significant when you consider the amount of money spent on Idle Compute. Min-Max Scaling (MMS) and Z-Score Normalization (ZSN) were the two main methods used to do data Preprocessing; IQR outlier detection was also used in this research. OOA-PSO was used for feature selection, and data Segments were created using Sliding Windows. The training used ResNet50 (transfer learning) with Adam optimizer and five-fold cross validation. The CNN-LSTM hybrid forecast approach combined with Deep Reinforcement Learning outperformed all of the other baseline algorithms in terms of Makespan, Energy, and Utilization.