Modeling the Time Dependent Behavior of Computer Systems
John Pfefferle, USDOE Office of Science (SC), US Atomic Energy Commission (AEC) · 1973
This study develops a procedure for modeling computer systems and for analyzing their performance. Here, computer system performance is measured by the actual or potential utilization of its components, irrespective of the effect on any particular job. The utilization of a component is related to the queue length at that component. A knowledge of the time varying behavior of the queue length is useful in explaining time averaged utilization measures. A computer program based on a queuing model was written to simulate the behavior of computer systems at the device level. The program is flexible enough to allow parametric variation of such things as the data channel structure, the scheduling methods used, and the service time distributions in force. The computer program also keeps a running time average of the utilization factor and mean and standard deviation of the queue length at each component of the simulated system. In addition, the lengths of the queues are sampled at equally spaced time intervals during the simulation. From the resulting time series the program constructs two measures of the time varying activity at each component: the histogram and the power spectrum. A series of nineteen experiments was run with the computer program. The experiments tested FCFS, SPTF, LFTF, RR, and preemptive SPTF scheduling and exponential, hyperexponential, hypoexponential, and constant service time distributions. The conclusions drawn from these experiments show the usefulness of time varying measures of activity and some specific phenomena associated with certain computer configurations or scheduling methods.