Introduction To Computer System Performance Evaluation

Rating:
Summary: Somewhat dated...but still has much to offer
Review: This book is now ten years old but it still has much to say regarding the mathematical and simulation modeling of computer systems. In addition, many of the results in the book are directly relevant to computer network modeling and simulation, the latter taking on enormous importance due to the explosion of the Internet and of ever-changing WAN technologies. Also, the author devotes a lot of time to the validation and calibration of performance models. This is a topic that is usually glossed over in other books on the subject.

Chapter 1 is an introduction to performance measures and evaluation techniques. It is out of date in some places, due to the hardware and operating systems the author discusses at various places in the book, these being available at the time at which the book was written. The SPEC benchmarking standard is discussed, but this benchmark has changed considerably since the book was written, and now has its own website. The author is careful to note the difficulties in applying performance evaluation measures used for single processor machines to those with highly parallel architectures.

In chapter 2 the author discusses hardware, software, and hybrid measurement techniques for measuring computer system performance. "Primitive" measurements are defined, and he explains how a condition for measurement can be recognized as sampled or trace monitoring. The pitfalls involved in performing measurements in these three techniques are discussed, along with other issues that effect measurement. One of the most interesting points the author makes here regards the use of the performance index and its connection with the control parameter.

The next chapter is very useful for those who need to do simulation modeling. Since such modeling is dependent on random number generation, the author discusses this along with the statistical data analysis techniques that must be used when analyzing the results of these simulations. He describes three different ways to organize simulation runs, and how to choose the input factors to make the simulations less time-consuming. And, regression analysis, so important to performance modeling, is treated in great detail, along with variance reduction techniques.

Queuing theory, of upmost importance in computer and network modeling, is discussed in chapter 4. The open and closed models are treated in detail, and the performance parameters of queues delineated. The throughput, utilization, and arrival laws, along with the famous Little's law are derived. The reader involved in Web server performance modeling will benefit greatly from the reading of this chapter. And, most importantly, the author discusses techniques for validating queuing models, a rare undertaking in books on this subject.

The dialog gets much more mathematical in chapter 5, which discusses the stochastic processes of isolated X/Y/c queues, for X the interarrival-time distribution, Y the service-time distribution, and c the number of servers. Steady-state analysis of several different queues is accomplished, and the response-time distribution is calculated for an M/M type of queuing system. Transform techniques are used to solve batch systems via the method of stages.

Product-form solutions are the subject of chapter 6, wherein the author considers both open and closed product-form queuing network models. The mean value analysi algorithm and a few others are discussed as exact algorithms for these kinds of queuing networks. Mean value analysis has arisen reently as an important algorithm for the study of wide-area network models. This and other algorithms for product-from networks are considered in much more detail in chapter 7, where convolution and the LBANC algorithms are also treated.

Approximation techniques for solving complex queuing network models are discussed in chapter 8. Flow-equivalent aggregation techniques, which allow one to collapse a subnetwork of a queuing network to a single station with a particular rate function are detailed. These techniques however, as the author points out, are very difficult to validate, and he does not discuss methods to do this. In addition, the author defines an interesting notion of a "conditional response time" for non-product-form scheduling disciplines.

Even more advanced mathematical techniques are brought in beginning in chapter 9, wherein the author first tackles the E(k)/M/1 queuing system using generating functions. Readers will need some background in complex variables to follow the treatment here. The all-important M/G/1 queuing system is discussed and the author calculates the response-time distribution and busy period for it. The latter is important in the modeling of token ring networks. Matrix analytic techniques are then used to analyze nonexponential queuing systems and queues in random environments. The author also introduces Markov modulated Poisson processes in this context. These have become very important recently in the modeling of the Internet.

Chapter 10 is more specialized, in that the author discusses techniques for find product-form solutions for networks with chain-dependent capacities, variable chain populations, rejection blocking, etc. The convolution and mean-value analysis algorithms are generalized to deal with these kinds of networks. Rejection blocking networks are interesting from the standpoint of TCP/IP networks, as they model a dropped packet that must be retransmitted by the client. Asymptotic expansions are used to analyze the behavior of normal usage networks. One example of this of current interest is a satellite network.

Chapter 11 moves away from solution methods and concentrates on performance bounds for queuing networks. This is done for both load-independent and load-dependent stations, and for multiple-chain networks.

In chapter 12, the author considers a very popular method for modeling the behavior of concurrent systems: Petri net techniques. These are briefly reviewed and then generalized to the stochastic case, followed by a discussion on how to apply them to the modeling of multiprocessor systems.

The last chapter of the book outlines various applications of the methods developed in the book, one of these being token-ring networks. Although these networks are becoming rare these days, some large businesses are still employing them and so an understanding of their behavior is still important. Performance optimization problems and the of modeling fault-tolerant systems are also considered.