Section 9.5. Perspective

---

9.5 Perspective

Performance modeling is a complicated subject. I hope that this chapter helps you understand the technology. But even more importantly, I hope that you can more understand the constraints of performance. I've met many performance analysts that pitted themselves in a losing battle against immutable laws of nature. I have constructed this chapter to help prevent you from falling into the same traps. For a reasonably complete final perspective on the chapter, I offer you the following points:

• Trial-and-error is an inherently inefficient, expensive, and unreliable performance optimization method. When a system meets the constraints required for using a mathematical model, it is much more efficient to base performance optimization decisions upon the model.

• Response time is virtually the only performance metric that end users care about. To the user , response time is the duration between the issuance of a request and the return of the first byte fulfilling that request. To the queueing theorist, response time equals service time plus queueing delay. We can reduce response time either by reducing service time, or by reducing queueing delay.

• On busy systems, response time degrades because of queueing. You can reduce queueing either by reducing workload, or by reducing service times. Performance analysts often forget that workload reduction is often a legitimate business option. A mathematical queueing model helps the analyst understand the economic tradeoffs required to meet both the functional and performance goals of a business.

• The M/M/ m queueing model is a well-researched, well- tested model for predicting performance of systems whose interarrival times and service rates are exponentially distributed. Many Oracle systems meet these criteria. This chapter provides a full Microsoft Excel implementation of M/M/ m , a Perl program to test whether a sample appears to be taken from an exponential distribution, and detailed instructions for using the model in an Oracle project.

• One of the most important virtues of using a queueing model is that it structures our thinking about response time. It reveals the concrete mathematical relationship among the parameters of workload, service rate, and expectations. Furthermore, it highlights the notion that the way to optimize the business value of a system is to consider all of these parameters to be negotiable.

• Our worked example showed a common business case: a scenario in which even though CPU is the system's bottleneck, adding CPU capacity doesn't help the analyst meet the system's performance requirements. The model in this example reveals what it often reveals in reality: that the most economically efficient way to improve the performance of a system is to eliminate unnecessary workload. The two principal ways to eliminate unnecessary workload are to avoid unnecessary business functions, and to reduce application code path lengths.

• The M/M/ m queueing model ignores a number of factors that the performance analyst must take into consideration. For example, the model assumes perfect scalability across all m service channels. The model is in several ways optimistic. An optimistic model that forecasts poor system performance is a proof that the modeled configuration will suffer poor performance in reality. However, an optimistic model that produces a positive verdict does not prove that the modeled configuration will perform well in reality; some un- modeled scalability barrier can still ruin the performance of your project.

---

Top