A method of resource management in cloud environments

Abstract

The article presents a comprehensive method of adaptive management of computing resources in cloud environments, aimed at increasing the efficiency of the use of infrastructures under conditions of dynamic and resource-intensive load. The proposed approach combines the mechanisms of hybrid time series forecasting using deep learning models (LSTM), statistical methods (ARIMA, Prophet) and multi-criteria optimization of the process of scaling and distributing tasks. Particular attention is paid to SLA-oriented classification of requests, which allows you to take into account the priority, criticality to latency, and computational intensity of tasks, increasing the level of quality assurance of service. The developed architecture of the method has a modular construction and provides for the integration of a predictive component, a context-oriented scheduler and a mechanism of adaptive load balancing that operates in real time. A feature is the use of a dynamic choice of a forecasting model depending on the characteristics of the load, which provides an increase in the accuracy of estimation of future resource needs. The task assignment algorithm implements multi-factor ranking of compute nodes based on latency, geographic proximity, load level, and power consumption, which helps to achieve a compromise between performance and economy. Practical confirmation of the effectiveness of the method was carried out by simulation modeling in the Kubernetes environment using real scenarios of variable query intensity. The experiments conducted demonstrated a significant reduction in the average response time and the number of SLA violations compared to basic reactive autoscaling strategies. Also, there was an improvement in the indicators of the level of resource use and a decrease in total energy consumption due to the flexible shutdown of excess capacities. The results of the study confirm the possibility of applying the proposed method in the conditions of highly dynamic traffic characteristic of service-oriented and edge-cloud architectures. Due to its ability to self-adapt and a hybrid approach to forecasting and resource management, the method can be effectively implemented in distributed computing systems that work with strict SLA requirements and increased requirements for energy efficiency. Based on the results obtained, the proposed solution is a promising basis for further research and practical use in the field of cloud technologies, IoT infrastructures and federated computing platforms.