Performance Evaluation of Operating Systems Using Approximate Analytical Methods

The motivation for the following paper were three models for a variant of a UNIX based multiprocessor operating system that was developed at the Institute for mathematical machines and dataprocessing IV [Jung 91]. These models were formerly analyzed by Markovian methods to get performance measures and the measurements of the real system were also available. This models had the following problems : Class switching is allowed. There are more job classes in the system available which can switch their class. The problem then is that the number of jobs in each class is not longer constant but depends on the time when you examine the system. Jobs have priorities with a mixed priority strategy. That means that some job classes can be preempted but others not. To reduce the time to analyze the system a new approximate technique is developed. The technique is based on the well known 'Mean Value Analysis' (MVA). This new technique can be applied to open and closed queueing networks. 1 Description of the System In this section three models of a UNIX-based multiprocessor operating systems [Jung91] are presented. For these models we developed a new technique to compute the performance measures as these models can not be analyzed with standard techniques. 1.1 System assumptions the network is closed and there are K jobs in the system. there are several job classes in the system and class switching is allowed. the service times are exponentially distributed. the peripheral device system consists of three magnetic tapes. the system has three job classes user, kern, driv with the priority order driv > kern > user. jobs in the context user can always be preempted by driv and kern jobs. Other preemptions are not possible. the system is in steady state.