A hybrid approach using attention bidirectional gated recurrent unit and weight-adaptive sparrow search optimization for cloud load balancing

Abstract

With the evolution of cloud computing (CC) technologies, there is a growing insistence for the maximum utilization of cloud resources, therefore increasing the computing power consumption of cloud’s systems. Cloud’s Virtual Machines (VMs) consolidation imparts a practical mechanism to minimize energy consumption of cloud Data Centers (DC). Efficient consolidation and migration of VM in the absence of infringing Service Level Agreement (SLA) can be arrived at by making decisions proactively based on cloud’s future workload prediction. Efficient load balancing, another major issue of CC also depends on accurate forecasting of resource usage. Cloud workload traces reveal both periodic and non-periodic patterns with the unexpected peak of load. As a result, it is very demanding for the prediction models to accurately anticipate future workload. This prompted us to propose a method called, Attention Bidirectional Gated and Weight-adaptive Sparrow Search Optimization (ABiG-WSSO) to accurately forecast future workload with minimal makespan and overhead. The proposed ABiG-WSSO method includes Attention Bidirectional Gated Recurrent Unit (ABiGRU) and Weight-adaptive Sparrow Search Optimization (WSSO). Attention Bidirectional Gated Recurrent Unit (ABiGRU) is initially designed that along with the use of Bidirectional Gated Recurrent Unit (BiGRU) and adaptation of attention mechanism aids in predicting future cloud load requirements accurately. Next, Weight-adaptive Sparrow Search Optimization (WSSO) algorithm is employed in fine-tuning the parameters of the ABiGRU model for accurate and optimal load balancing performance. The WSSO algorithm is applied to optimize ABiGRU model hyperparameters (i.e. learning rate), to enhance its prediction accuracy. Comprehensive simulations are carried out using the gwa-bitbrains dataset to verify the efficiency of the proposed ABiG-WSSO method in boosting the distribution of resources and cloud load balancing. The proposed method achieves comparatively better results in terms of better makespan time, energy consumption, associated overhead and throughput.