Performance analysis of computer networks
Steven Woolet, Kishor Shridharbhai Trivedi · 1993
As computer systems and networks continue to grow in size and complexity, the techniques and tools used to solve models of these systems must be capable of giving answers to the old questions as well as new ones. Models which give strictly performance measures or only dependability measures are not enough. Mean value results often do not give the breadth of information required. This thesis describes some modeling methods which can be used to meet the requirement for obtaining new and better answers. The Markov reward model has been shown to be useful in combining performance and dependability. We show new methods which can be used with this model to gain better insight into system performability. We also show a method which uses response-time distributions as a reward rate for such a model. System users are typically very concerned about the amount of time required for the system to respond to their input. Mean values of the response time may not give enough information. However, the ability to solve for the response-time distribution of a system being modeled is not always straight-forward. We describe a method of approximating the response-time distribution for queueing networks using Markov chains. In some cases the results are exact. Techniques which allow the modeler to consider computer networks under varying load conditions through the use of transient measures are described. If the variations in load are periodic in nature, a method is shown to reduce the number of calculations required in obtaining results for extended periods of time. For the modeling of practical computer networks, many times there are large numbers of items to consider (workstations, packets, etc.). This can lead to the generation of prohibitively large Markov chains to be solved. A computer network model is described that is solved using decomposition methods to reduce state-space sizes. Each of the different techniques is illustrated by an example. Three different types of examples are used; fault-tolerant computer systems, real-time computer systems, and computer networks. This shows that advanced methods are applicable and useful for many different types of models.