Browse Books

Reviewer: Pieter S Kritzinger

Here, for the first time, we have a book on computer performance analysis which does not confirm the worst fear of the practical analyst: namely, that queueing network models are for theoreticians only. The book is written by authors who are all well known for their theoretical research, but who obviously have enough experience in the field to be aware of the practical issues. The book is concerned exclusively with computer system analysis using queueing network models. Although it provides new insights and different viewpoints for even the most experienced computer performance analyst, it should not be recommended as a university textbook beyond a fundamental course in computer performance evaluation. There is too much of a cookbook flavor for it to be an advanced reference text. Its strength as a practical handbook lies in this very fact, since the methodology of queueing network modeling must be learned through practice and cannot be absorbed through a course in queueing theory. This is the main theme of the book which, accordingly, is full of practical and realistic case studies. Rather than just providing the methodology, however, the underlying theoretical assumptions are always clearly explained using arguments and terminology which should be clear to persons not necessarily knowledgeable about queueing theory. The book is divided into five parts. The first four chapters set the stage in that they take the reader through the various logical steps required when conducting a modeling study. Also presented are the theoretical arguments in a popular version of stochastic queueing theory known as operational analysis. While clearly enthusiastic about the applicability of queueing networks, the authors do point out that in the very simplicity of these models, their danger also lies; one should always be clear about the assumptions being made, the motivations for their introduction, and the arguments for their plausibility. Furthermore, the reader is made aware of the fact that queueing networks cannot model exactly certain very important characteristics of real systems such as blocking, simultaneous resource possession, and priorities. The next four chapters deal with the general analytic methods, ranging from techniques for computing performance bounds quickly for those systems for which the real workload is not known to models with multiple-workload classes. While presenting the theoretical foundations and the latest algorithms for solving the models, the authors are sufficiently realistic to point out that the major advantage of multi-class over single-class models (i.e., the representation of more workload detail) is also a major disadvantage: it is difficult to find the detailed input parameters in practice. The last of the four chapters in the second part of the book deals with hierarchical or flow equivalence modeling, which is the process of splitting one model into a number of smaller submodels, each of which is then analyzed in isolation. The third part of the book deals with those subsystems of a computer system which the analyst may wish to represent in greater detail than would be possible with the class of queueing networks introduced up to that point in the book. However, the chapter on modeling memory restraints, including swapping and paging, is disappointing. This is not surprising since accounting for these important effects in a model is very difficult and no satisfactory method which works well in all cases is known. Also discussed in this part of the book is the modeling of disk subsystems, with or without RPS. The authors deal with this subject rather comprehensively and, for the sake of completeness, should have devoted more than a brief mention near the end of the book to ways of incorporating tape devices into a model. The modeling of systems with multiple processors, or with processors using disciplines such as priority scheduling, is dealt with in some detail. Since load-dependent service centers, used to model closely-coupled multiprocessors and flow-equivalent centers, are central to many models in the book, the authors might have done better to introduce the theoretical aspects in the main body of the text rather than leaving it to the last chapter. The latter may inadvertently create the wrong impression with the general reader that, although load-dependent centers as such are theoretically possible, they do not have much application in queueing network models. The second to the last part of the book deals in some detail with the question of mapping the real or projected system and its real or projected workload onto the queueing network model. This process, which the authors call parameterization, is central to modeling building. The results of the performance projection study can be no more accurate than the parameter values provided. This part, therefore, deals with the important aspect of where to find the data and monitoring in general (when to monitor__?__ for how long__?__). Also discussed are ways to choose workload classes and to generally choose parameter values for systems which exist, are still evolving, or are only just being proposed. The last part of the book discusses the modeling of systems which do not have a central server of some sort, such as data communication networks. The discussion is somewhat superficial and unconvincing, although the authors do refer the reader to the original publications for more details. Equally, the authors, while stressing the advantages of commercially available queueing network software, never really discuss the use of queueing network modeling software packages in general, as the dustcover claims. The reader expecting a catalog of currently available software packages and the modeling facilities they offer will be disappointed. Such readers may want to refer to the Proceedings of the Conference on Modeling Techniques and Tools for Performance Analysis held in Paris in May of this year [1]. Finally, as is so often the case with textbooks, it is a pity that the authors do not supply answers to the many exercises included. It should be clear from this review that this is a good, practical book which is hard to fault. It is strongly recommended for the professional computer-performance practitioner. Since the book is packed with practical advice, the authors could have added another warning: namely that despite the most convincing analytical performance-prediction study, the final decision about a major computer-system expenditure is nevertheless made in the company boardroom based upon issues which all too frequently bear little relationship to the results of the technical study.