Description
This dissertation studies the problem of designing the configuration of local area network-based distributed systems by a performance modeling approach. The study concentrates primarily on the interactive transaction-oriented computer systems connected by an Ethernet-like network. Major transaction types are chosen to represent the workload; each type is characterized by the demands on various computing resources. A two-step methodology is described that produces the initial configuration; this configuration is the one to which more detailed queueing network models are applied. Based on the queueing model results, the methodology then iterates to refine the configuration of the distributed system.
Various transaction types are first assigned to host systems to balance CPU utilizations. We then distribute the shared files among host systems to minimize total remote file accesses. Queueing network models for our distributed systems are constructed from a set of submodels of the host systems, of the local area network, and of the file servers if any. Model parameters are derived from our workload data, measured in an interactive transaction-oriented business system. Two examples are provided to show how the configuration of a local area network-based distributed system can be designed by using our workload data.
Attempts are next made to capture in our queuing network models the difference in resource demands during the preparation phase and the execution phase of a typical transaction. As a result, we introduce the special class of phase-free queueing network. It is shown that a general product-form queueing network can be reduced to an equivalent phase-free product-form queueing network. It is also shown that the per-class throughputs, mean queueing time, and mean queue length in the original network can be calculated from the values of the aggregate indices of the phase-free network.
Lastly, a special class of local area network-based distributed systems, that of the workstation-based systems, is examined in some detail. Workload clustering is proposed as a method for reducing the number of chains in the model to make it more tractable mathematically. File server design issues are investigated, and design guidelines are recommended based on workload data and performance goals.