Recitation: Generating Random Variables for Simulation

• Usually we don’t have enough trace data to collect meaningful system measurements (statistically significant ones). • Sometimes we might purposely want to generate a Poisson arrival process, so that we can study effects of other variables. • If you’ve used empirical data to create closed-form analytic curve-fit, that analytic distribution can be very accurate. For example: the SURGE Web workload simulator was created by studying millions of logs of Web requests at various Web sites. Analytical distributions were determined which fit the distribution of service demands, the file popularity, the “thinktime” of clients, etc.. The SURGE simulator now generates “hypothetical” requests from these distributions and is a useful performance prediction tool.