Modelling with queues: an empirical study

We review the principles of computer performance modelling with queues, and validate the models using the empirical data. We focus on modelling a special configuration of two computers working in tandem: a mirroring system. Distributed system under investigation has implemented as a series of Java applets communicating using the functions of java net.* package. The mirroring aspect of the system is modelled analytically with a difference queue. The difference queue is defined for a network of two queues in parallel, as the queue consisting of elements present in one queue and absent from the other queue. The queueing models under investigation are M/M/1, and M/D/1 queues. In our study we have observed that the service times were not completely random on a typical server and could not be approximated well by exponential distribution. Our results have shown that the service was almost (i.e. with a very small variance) deterministic. We have also found that the d-queue model approximates well the behaviour of the mirroring system under low load conditions, and also provides an upper bound on the delays under high load conditions