Omelchenko V. Information technology for computing resources management in Kubernetes cluster

The dissertation work is devoted to solving the problem of automating the management of computing resources in Kubernetes clusters. Computing and financial resources are limited, which requires a constant search for a balance between the level of service quality and the amount of resources. Automation of management processes allows for an increase in the efficiency of using the cluster's computing resources while maintaining the required level of quality of services provided. The existing forecasting methods are analyzed for the feasibility of their use for proactive scaling. Experimental studies of the accuracy of the selected methods were conducted on typical workload templates based on historical data of various length. Based on the experiments, a set of methods was selected for predicting workloads in proactive scaling. A hybrid forecasting method with a combination of long-term and short-term forecasting components is proposed. An experimental study of the proposed method was conducted. The results demonstrate an increase in overall accuracy from 91% to 95% on the test data, provided that there are anomalous patterns in the historical data. A method for proactive scaling of computing resources is proposed. An architecture that includes data, forecasting, application, and decision-making modules is proposed. An approach for the method to work in the absence of data or low accuracy of the obtained forecasts is describedA software module for horizontal proactive scaling was implemented in Kubernetes using built-in tools. An experimental study of the developed software module on a Kubernetes cluster was conducted to compare it with static and reactive scaling approaches. Compared to the redundant static approach, a similar level of service quality was obtained using 46% less resources. Compared to the reactive approach, the average application response time decreased from 160 ms to 23 ms. A hybrid scaling method is proposed that includes reactive and proactive components. The combination of these methods allows the use of proactive control in the presence of accurate forecasts. An indicator of the transition between proactive and reactive control is proposed based on comparing the compliance of the obtained forecasts with the current load on the application. The level of accuracy of forecasts is determined within a given number of iterations. A software module was implemented using the developed proactive method and a built-in solution for reactive scaling in Kubernetes. An experimental study of the developed program module was carried out. The results demonstrate the ability of the proposed method to identify anomalies and transfer control between components according to certain rules. A hybrid scaling method is proposed that allows coordinating the work of horizontal and vertical scaling components. A coordination module is proposed to coordinate the configuration of the vertical and horizontal components. Based on the proposed method, a program module for working in Kubernetes was developed. The results of the experiments demonstrate a 65% reduction in unprofitable reservation of computing resources compared to the static approach on synthetic data. The information technology for managing computing resources in a cluster is described. The decomposition of information systems into separate modules is proposed. The functionality of each module and communication between modules and the cluster are described. Based on the proposed architecture, the implementation of information technology in Kubernetes clusters is described.