PERFORMANCE-ORIENTED DESIGN OF DISTRIBUTED SYSTEMS
Abstract (summary)
Existing methods support one of two approaches to the design of distributed, high-performance architectures. Queueing models support performance evaluation at a very abstract level of the design. Logical design models support evaluation of correctness and realizability at lower levels in the design process. This dissertation defines a modeling framework in which performance and correctness can be studied. The UCLA Graph Model of Behavior is selected and extended to support integrated Performance-Oriented Design. The GMB supports formal Control Flow Analysis, similar to the Petri net, which is useful for detecting the potential for deadlock, the potential for unbounded demands on system resources, and other control flow anomalies. Extensions defined and implemented in this work are proven to be consistent with the control flow analysis formalism. The proposed extensions include the capability to specify finite-capacity service centers, pre-emptive and non-preemptive service disciplines, data domain queues which support customer classes, deterministic customer routing, and under certain restrictions, priority service. The resulting framework is applied to evaluate the performance of a ring network proposed for distributed execution of functional programs. Issues in modeling of low-level hardware and concurrent software which can be partially solved by the above extensions are also discussed, leading to better support for realization of graph models.