In the thesis, the important scientific and practical problem of increasing the energy efficiency and performance of workload processing in the information and communication network (ICN) while meeting the requirements for the availability of the workload processing system was solved. The requirements’ analysis conducted for each of these workload types in accordance with the recommendations of the International Telecommunication Union allowed determining the main performance indicators of the distributed workload processing system as part of the ICN and the server cluster as a unit of the distributed data center as part of the ICN, in particular: energy efficiency and performance indicators of workload processing, as well as the system availability factor. Based on these indicators, an optimality criterion of workload processing in ICN was proposed. In order to systematize and formalize the workload processing process in the information and communication network as an object of research, an ontological model of a distributed workload processing system was built. In order to obtain a quantitative assessment of the relationships between the defined efficiency indicators and the parameters that affect them, a mathematical model of the distributed workload processing system within the ICN as a queuing system (QS) was built. While building the model, a method of transition from a non-stationary non-ordinary input requests flow to a stationary ordinary flow was proposed by discretizing the intensity curve of the input workload and using the transition to sets of servers, which made it possible to significantly simplify calculations with permissible losses of model accuracy. For the discretization of the input workload arrival rate curve, the use of the quantization by levels was proposed, which made it possible to match the size of the discretization step with the rate of change of the input workload arrival rate. To determine the quantization step, a method of calculating threshold values of input workload arrival rate as a function of the number of computing nodes in the system is proposed. Based on the constructed mathematical model, a method for calculating horizontal scaling patterns is proposed, which allows determining the optimal number of active computing nodes in the system at each time interval, which is determined by the rate of the input workload arrival rate change. The methods of determining individual energy consumption models of computing nodes were analyzed and the expediency of their use in the workload processing process in ICN was substantiated. The built mathematical model of the system in the form of QS and the considered methods of determining individual energy consumption models of computing nodes became the basis of a new comprehensive method of energy-efficient workload processing in computing nodes of distributed data centers. The proposed comprehensive method differs from known ones in the use of individual models of computing nodes’ energy consumption, a combination of the advantages of horizontal scaling approaches and energy-efficient scheduling, while taking into account dynamic changes in the input workload arrival rate, which made it possible to increase the energy efficiency of the workload processing without loss of performance and subject to compliance with system availability requirements. As part of the proposed comprehensive method, the existing approaches to horizontal scaling of the computer system were improved by using individual models of computer nodes’ energy consumption and mechanism for predicting dynamic deviations of the input workload arrival rate, which made it possible to ensure more intensive use of the most energy-efficient equipment and to respond in time to unpredictable changes in the input workload arrival rate.
On the basis of the proposed comprehensive method of energy-efficient workload processing, software for managing computing resources has been created, which allows increasing the energy efficiency and performance of distributed workload processing while complying with the requirements for system availability, and can be used to increase the energy efficiency and performance of workload processing in edge and central cloud within the 5G network architecture. The effectiveness of the proposed comprehensive method and the software based on it was verified using the methods of laboratory experiment and simulation modeling. The performance of the proposed comprehensive method in comparison with the known Backfill and Round Robin approaches in terms of energy efficiency was 9.953% and 26.382%, respectively. The performance gain was 5.593% and 49.458% respectively. At the same time, the proposed comprehensive method ensures the fulfillment of the requirements regarding the system availability and gives a gain according to the proposed optimality criterion by 15.722% in comparison with Backfill and by 88.887% in comparison with Round Robin.
