A Mathematical Model of Load Distribution for Data Center Server Clusters

Abstract

An approach to estimate the performance of a load distribution and balancing system for data center server clusters is proposed, which takes into account the current load (occupancy) of servers. This system is described in mathematical terms, focusing on the server states at each instant of receiving service requests in a steady-state mode. The capacity and response time to customer requests are used as the main performance indicators. These indicators are estimated using queuing theory and simulation modeling as mathematical apparatus. The flows of service requests are stationary and ordinary. The receipt and processing of service requests are described by Markov processes. The load distribution and balancing system is represented as a multi-channel queueing system with a limited queue length, which includes an unlimited buffer (disk memory) for all servers in the cluster and input buffers of limited capacity for each server. The effectiveness of the proposed model and the theoretical results are verified by simulation, and experimental data are obtained.