Selecting and generating variates for modeling service times

Abstract

We have tabulated the computer times required to generate 200,000 samples from 7 different distribution types using implementation of the INSIGHT simulation language on 4 different computer configurations. This study was motivated by our recent migration from using triangular to using Johnson (S B) input distribution when building simulation models in a data-poor environment. We were concerned that the increased flexibility of the Johnson (S B) might be obtained at the cost of greatly increased computer run times, but no such trade-off was experienced. The variate-generation times for Johnson (S B), Weibull, and lognormal variates were found to be similar—somewhat slower than for uniform or triangular variates, but substantially faster than for gamma or beta variates. On an 80386-based personal computer the differences in speed between beta and gamma variates and all of the other standard distributions were magnified when the 80387 math copressor was disabled. As another part of this study, we obtained profiles of large-scale health-care simulation experiments and other real examples that were executed on a mainframe computer (Gould NP-1). The results showed that variate generation rarely exceeded 3% of the total run time.