A Load Balancing Method for a Data Center Computing Cluster

Abstract

This paper describes a load balancing method for a data center computing cluster. The method involves a probabilistic approach to the proactive forecasting of packet traffic states based on statistical, nonlinear, and spectral analysis. Due to the fractal properties of network traffic, it is possible to forecast reliably the occurrence of traffic bursts and declines on separate time intervals and identify periods with possible overloads on servers and network equipment. Therefore, it is possible to develop methods for the effective planning and distribution of tasks within data centers to ensure the statistically uniform loading of their functional elements. The spectral analysis of the time series is performed using the normalized deviations of actual levels from the smoothed ones. No significant peaks in the spectral estimates mean no periodic fluctuations. As shown below, summing the cycles of different-period dynamics of the time series for the most significant harmonics of the spectrum determines the moments of occurrence of the subsequent anomalies. The most significant harmonics of the spectrum are identified by studying its spectral power density using the Fourier transform. The developed method is a solution for the efficient planning and distribution of tasks in a data center computing cluster that optimizes the use of resources, accelerates task execution, and reduces application processing costs.