Analysis of a complex retrial queue: Resource allocation and system performance in a dial-up modem pool

Internet Service Providers, universities, and businesses have long provided remote access to central computer networks, via telephone lines, through collections of modems called dial-up modem pools (DMP). Recent strong growth in the popularity and usage of the Internet has resulted in excess demand for service in many DMP systems. When system operators can not add capacity, due either to time or funding constraints, they must manage their limited resources to provide all users fair and equal access to the system. Imposing time limits on each connection to the system is a common method of dealing with this problem, but little is known about how the system will react to these limits. This research examines the relationship between such limits and pertinent system performance measures, and whether these relationships can be modeled for use in predicting future system performance.; Dial-up modem pools fit a general class of queuing systems called retrial queues. Direct analytical results in such systems are rare, and none of the existing literature addresses any model similar to a DMP system. As such, simulation is used to generate system performance measures given different system configurations and customer behavior patterns. Real-life data from Indiana University DMP systems are used, when available, for simulation modeling. Results from the simulation experiments show that time limits have a definite and nonlinear effect on all of the performance measures. It is also shown that for two key and offsetting performance measures, time limits can have a net positive effect on customer service over a wide range of such limits.; Linear regression and neural networks models are developed to explore whether these systems can be modeled well enough to accurately predict future system performance. These models are tested with unseen data, and predictive statistics are computed. Performance of the models varies across system performance measures, but the neural network models generally perform best, especially on the test data set.