We assume that the work a job has to execute, , is drawn from a Hypergeometric distribution with mean and coefficient of variation (1, 5, and 30 are considered). This is consistent with variations in service demands used in previous studies [5, 26] and those observed at one supercomputer installation [9]. We also assume that there is no correlation between the amount of work a job executes and the efficiency with which it is executed. Although this might not be true for all applications, we want to separately examine the effects of service demand and efficiency on the mean response time obtained with different allocation policies. This is not possible if efficiency and work are correlated.
We model the fact that jobs execute with different efficiency by using the execution rate function, F, for all jobs and choosing uniformly between and . This distribution is similar to that used in previous studies [23, 30] except that we ensure that the distribution is uniformly distributed in rather than . We believe that this is what was actually intended in the previous studies.
Each workload executes M jobs, whose arrival follows a Poisson distribution. Each experiment is repeated a number of times using different random seeds in order to compute confidence intervals. The number of jobs and repetitions used for each experiment was chosen in order to achieve 90% confidence intervals that are within 5% of the mean. We use the bootstrap method for computing confidence intervals since it is robust for small numbers of repetitions and for non-normal distributions of observed means [34].