Book contents
- Frontmatter
- Contents
- Preface
- Acknowledgements
- 1 Introduction
- 2 Metrics of performance
- 3 Average performance and variability
- 4 Errors in experimental measurements
- 5 Comparing alternatives
- 6 Measurement tools and techniques
- 7 Benchmark programs
- 8 Linear-regression models
- 9 The design of experiments
- 10 Simulation and random-number generation
- 11 Queueing analysis
- Appendix A Glossary
- Appendix B Some useful probability distributions
- Appendix C Selected statistical tables
- Index
10 - Simulation and random-number generation
Published online by Cambridge University Press: 15 December 2009
- Frontmatter
- Contents
- Preface
- Acknowledgements
- 1 Introduction
- 2 Metrics of performance
- 3 Average performance and variability
- 4 Errors in experimental measurements
- 5 Comparing alternatives
- 6 Measurement tools and techniques
- 7 Benchmark programs
- 8 Linear-regression models
- 9 The design of experiments
- 10 Simulation and random-number generation
- 11 Queueing analysis
- Appendix A Glossary
- Appendix B Some useful probability distributions
- Appendix C Selected statistical tables
- Index
Summary
‘Do not plan a bridge's capacity by counting the number of people who swim across the river today.’
UnknownOftentimes we wish to predict some aspect of the performance of a computer system before it is actually built. Since the real machine does not yet exist, we obviously cannot measure its performance directly. Instead, the best we can do is to simulate the important aspects of the system. We then try to extrapolate from these simulations information about how the system will behave once it is actually built. Simulation may also be appropriate when we want to investigate some aspect of a system's performance that we cannot easily measure directly or indirectly.
When the system does not yet exist, there are many assumptions that must be made about the, perhaps not completely defined, system before it can be simulated. Simplifying assumptions are also necessary when simulating an existing system since it would most likely be impossible to simulate every small detail. If these assumptions are not realistic, the simulation results will not accurately predict how the system will ultimately perform.
Simulation has the advantage of being much less expensive than actually building a machine. Additionally, it is much more flexible than measuring the performance of a real machine. In a simulated system, we can quickly change important parameters that would be difficult or impossible to change in a real system, such as the size or associativity of the cache, for instance, to determine how the system's performance will be affected.
- Type
- Chapter
- Information
- Measuring Computer PerformanceA Practitioner's Guide, pp. 181 - 216Publisher: Cambridge University PressPrint publication year: 2000