The characterization and prediction of the behavior of a computer system is one of the main goals in any performance evaluation study. Moreover, if one is interested in comparing different computer installations, which may have diverse user communities and hence different natural workloads, workload estimation and comparison, as well as user utilization, become central issues of discussion.

In this paper we present the results of using a portable workload estimation technique, based on the terminal probe method, to characterize the natural workload of interactive computer installations and predict selected performance indices of interest. Studies performed on 12 computer systems show that there are system independent statistical characteristics of the observed performance indices which allow us to model system behavior and to compare installations. These results can be used as a basis for load balancing strategies in distributed systems. We have found a family of statistical models which fit our measured data in a comprehensive way; not only the mean and the variance of our distributions are well approximated but the modelled distributions fit the observed ones. Thus, order statistics can also be obtained.

All of the observed systems operate under the UNIX operating system. The methodology utilized is easy to understand, quick to implement, and may be applied without bringing the systems down. The more than 16,000 measurements used for this study were obtained over a span of two and a half years. Detailed analysis done with these data enable us better to control the duration of the data gathering period and, what is more, achieve predetermined statistical confidence in the modelled distributions.




Download Full History