G. Bolch, S. Greiner, K. Trevedi
Abstract: Performance evaluation of current computer- and operating systems becomes more and more important because of the growing system complexity. Therefore it is necessary to have a mathematical model of the complex system. A very well known and popular technique is modelling by using queueing networks because they are easy to understand and use compared to other modeling techniques. For so called product form queueing networks exist efficient exact and approximate analysis algorithms but most queueing networks of realistic systems do not fulfill the requirements of product form queueing networks. To solve this kind of networks Markovian analysis or discrete event simulation is used. The disadvantage of these techniques is that they are very time consuming to prepare, implement and run, especially when we want to analyze the system on a wide range of parameter values. Because of these disadvantages the goal must be to transform queueing networks, that can not be solved with standard analysis techniques, into networks that can be solved with standard analysis techniques (like MVA, SCAT,....). Problems that appear very often in realistic systems are : - Priorities with and without preemption. - Mixed priorities. - Class switching. - Heterogenous nodes. This problems appear especially in the field of operating system modeling. In this paper we show how to transform a given non product form queueing network (with the problems mentioned above) so that the network can be analyzed with a standard analysis technique. To demonstrate how this new technique works, we use a simple non product form queueing network and three models of a UNIX based multiprocessor operating system.
[Full Paper (ps,http) , 160 kB][Full Paper (pdf) , 300 kB]