UDC 004.75
LOAD FORECASTING USING RECURRENT NEURAL NETWORKS TO DETERMINE NODE PRIORITY IN A DISTRIBUTED SYSTEM
T. E. Zeynally, postgraduate student, assistant at the Department of Computer Science and Information Technologies, Moscow Polytechnic University, Moscow, Russia;
orcid.org/0000-0001-6462-9607, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
D. G. Demidov, Ph.D. (Tech.) Dean of IT Faculty, Moscow Polytechnic University, Moscow, Russia; orcid.org/0000-0002-2462-936X, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
This paper examines distributed systems composed of peer nodes that iteratively perform useful workloads. The specific feature of the systems considered is that priority is calculated among peer nodes – without a coordinator node. The problem of prioritizing nodes for executing specific tasks is defined for these systems, and existing scientific research on this topic is analyzed. The aim of the work is to develop a method for predicting the priority of a node for executing a payload in systems with the above-described properties. An analysis of iterative workloads characteristics is conducted to determine and describe possible prioritization scenarios and to identify parameters and characteristics that influence priority. The result of the analysis is a method for calculating priority that enables the system to adapt to current conditions and distribute the load rationally. The identified characteristics allow adaptation of prioritization scenarios for specific tasks and improvement of node resource utilization. The paper describes experimental setup – a system of three nodes iteratively performing useful workloads. The data collected from this setup is used to train a neural network. Methods for predicting node priority using machine learning techniques are described. The main result of the work is a recurrent neural network model that predicts the priority of a node for executing useful workloads, based on the information from previous time intervals. The paper describes model input and output data. The inputs to the model include characteristics of iterative process and time series of historical load indicators on a node. The output value is a priority indicator for next time interval. The accuracy of forecasting model is evaluated. The application of the proposed method helps to effectively utilize the resources of distributed information systems. Practical applications of the results obtained in the course of the work are presented.
Key words: : distributed system, information system, computer software, node prioritization, RNN, LSTM, machine learning, forecasting.