Predictable Scaling Behaviour in the Data Centre with Multiple Application Servers

Load sharing in the data centre is an essential strategy for meeting service levels in high volume and high availability services. We investigate the accuracy with which simple, classical queueing models can predict the scaling behaviour of server capacity in an environment of both homogeneous and inhomogeneous hardware, using known traffic patterns as input. We measure the performance of three commonly used load sharing algorithms and show that the simple queueing models underestimate performance needs significantly at high load. Load sharing based on real-time network monitoring performs worst on average. The work has implications for the accuracy of Quality of Service estimates.