Workload Data

 

There are three main approaches to providing a workload description for the simulation:

• Stochastic models of the workload have the advantage of being derived from probability distributions and are therefore statistically well defined. Stochastic modelling is a method often introduced for dealing with uncertainty (when there is variation but no discernible pattern), Franta [Fra77]. The main advantage of this approach is that parameters can be altered to provide widely different test situations. The simplifying assumptions used in creating a stochastic model can be difficult to defend, yet this is a popular approach.
• Traces of jobs started on a real system are recorded and used to drive the simulator. Provided a true cross-section of activity is recorded, this will provide the most realistic data. In addition, traces can be rerun for repeatability, and different traces can be used to study different workload situations. The drawback of this approach is that it is still a sample, and the data may only reflect the activity on the system from which it was recorded. Another problem is that obtaining detailed trace data from a real system, may interfere with the activities being recorded, thus reducing the value of the measurements.
• Synthetic mixes are obtained by observing the system under study, deriving rules and relationships to create a facsimile of the real situation. Like a simulator this approach has the advantage that it is under complete control and can be used to show effects on demand (i.e., rare situations can be created and the implications for the system can be studied). Care must be taken when constructing the models for a synthetic workload, as any results are open to challenge if the models do not represent a realistic workload.

All of the above methods require a description of a workload for which results can be obtained and analysed. Experiments should be repeatable and poor workload descriptions must be avoided on the principle of `garbage in, garbage out'.

With this in mind, the process mix hypothesis is most effectively tested with a real workload composition, as it is important that the proportions of the resource bound processes are realistic. A synthetic or stochastic model could easily misrepresent the system load, and lead to faulty conclusions. Therefore the most reliable choice is to use system traces to drive the testbed simulator.