ABSTRACT
In queuing theory literature, many models describe the dynamic behavior of computer systems. Good sources of such information (e.g. [5,7]) usually include stochastic models based on birth-death formulas, the convolution algorithm [2], load independent and load dependent mean value analysis (MVA) models [10], and BCMP networks [1]. Theoretical queuing models presented in computer literature easily explain phenomena such as bottlenecks, saturation, resource utilization, etc.; however, it is very difficult to find sources that show a second level of modeling, which focuses on the ability of models to also achieve good numerical accuracy when modeling real computer systems running real workloads. Although the phenomenology is important in the classroom, it is the numerical accuracy that counts in engineering practice. The usual task of performance analysts is to measure system performance and then derive models that can describe and predict the behavior of analyzed systems with reasonable accuracy. Those who try to model the dynamic behavior of real computer systems running real workloads frequently find this to be a difficult task.
