IBM
Contents Index Previous Next



A Performance Simulation Project


To develop a performance simulation and use it to obtain estimates about a particular system involves a number of steps. The SDL Suite may be helpful in many, but not all, of these steps.

The major activities in a performance simulation project are:

  1. Collect information about the behavior of the system to be simulated and define the purpose of the simulation, that is, what estimates should be the result of the project?
  2. Create a performance model for the system. This is often expressed as a queuing network. As the behavior of a system is usually too complex to be simulated in all details, simplifications have to be made.
    This modeling phase is the most critical phase in the complete project. A good performance model leads to relevant results, while a bad model leads to nonsense. The key question is what simplifications and abstractions you may make and still obtain relevant results from simulations of the model.
  3. Describe the performance model in detail in SDL and simulate it, using the library Simulation, until the SDL model behaves satisfactorily.
  4. Execute the generated simulation a number of times to collect statistical data from the model.
    You will normally need 10,000 to 50,000 samples (of for example a queue length or a waiting time) to obtain any significance in the data. This means that program executions will be fairly long. There are two ways to execute the performance simulation:
    • Use the library Simulation and start by performing the monitor commands "Set-Trace 0" and "Go", or
    • Use the library PerformanceSimulation. It is specially designed to execute performance simulations and does not include the monitor system. It will execute the performance simulation in the order 10 to 15 times faster than executing it using the library Simulation.
  5. During the execution of the performance simulation, the best way to handle data measured in the system is to write the data on file. You may then analyze the data files using packages or programs for statistical analysis to obtain, for example, mean value, variance and confidence intervals for these estimates. (The SDL Suite does not provide means for statistical analysis.)
  6. To validate these estimates it is common practice to compare the simulation results with results from mathematical methods applied on the system (simplified versions of the system). Queuing theory and the theory for queuing networks are the most relevant mathematical methods for such activities.

Some of these steps will discussed in more detail in the following subsections. For other aspects please see literature about performance simulations.


http://www.ibm.com/rational
Contents Index Previous Next