Trace-driven system modeling

This paper has attempted to introduce the basic concept of trace-driven modeling. As in all simulation approaches, one major concern is the precision of predicitive capability. The trace-driven approach lends itself to accurate predictive evaluations by successive calibration steps. Given an existing configuration capable of executing a given job mix in an operating-system environment, the time required to run the jobs can be measured and firmly established. On the other hand, the job profiles for this set of jobs can also be obtained using a trace program under the same conditions. The output from the model that runs with the parameters representing the physical system can be compared directly with the measured results. Because of this inherent property, accuracy of the model can be established prior to its use on a configuration that cannot be measured. In this paper, the model was presented as a demonstration of trace-driven modeling. Although no formal effort was conducted toward validation of the model, results from a preliminary investigation were satisfactory. The experience gained from this experiment has convinced us that the trace-driven approach is feasible for computing-system evaluation. In computing science, as in other scientific areas, changes generally occur in an evolutionary fashion. Therefore, the trace-driven technique may well be applicable for future system evaluations.