The purpose of the article is to improve the performance of cloud data center cluster servers by optimizing the distribution of computing load. Based on models of nonlinear dynamics, a method for predicting the states of fractal network traffic has been developed, which includes the stages of nonlinear analysis and linear forecasting and is the basis for a dynamic algorithm for load distribution between computing resources. The dynamic method of predicting data center load states is based on a mapping system presented in the form of an algebraic polynomial and a resource allocation matrix. The nonlinear determinism of the input load of the data center is determined by the assessment of Lyapunov exponents using the TISEAN 3.0.1 platform. The Takens-Manet theorem is used as the basis for the reconstruction of the phase portrait of the input load process. To eliminate noise and random disturbances in the numerical series of network traffic, the SSA singular spectral analysis method was used. The numerical value of the load deviation from the nominal level is used as a criterion for the effectiveness of the proposed solution. As a result of the conducted research, a dynamic algorithm for load distribution between the servers of a data center cluster has been developed to ensure a uniform load of the information system in conditions of fractal network traffic. The results of the numerical experiment showed that the use of the developed nonlinear algorithm can significantly improve the quality of distribution, load balancing and data center user service.