Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Electrical and Computer Engineering

First Advisor

Ahmed El-Amawy


The dissertation introduces a new class of bus-based systems called the Segmented Multiple Bus System (SMBS). One of the unique characteristics of the SMBS is that it allows the exploitation of memory reference locality even though it is a bus-based nondirect network. This comes from the architectural interconnection feature that the SMBS can be viewed as a large-scale multiple bus system (MBS) that has been partitioned into smaller partitions called segments. Each such segment is in effect a small conventional MBS whose size is chosen so as to avoid bus loading problems. The SMBS overcomes the architectural limitations of bus-based shared memory systems while maintaining their advantages in terms of high degree of fault tolerance, ease of expansion and ease of programming. In addition SMBS's are scalable; unlike conventional MBS's. Another interesting feature of the SMBS is that it supports wormhole routing which is traditionally used in direct network topologies. We develop performance models to study the SMBS with wormhole routing. Ours is the first attempt to adapt wormhole routing to a bus-based nondirect network. In our performance modeling, features of both direct and nondirect networks are incorporated. We include the effect of blocking and pipelining properties of wormhole routing in the analysis. The bus group of each segment is modeled as a flow equivalent service center which represents a load dependent service center. Two performance models, one assuming single flit buffers and the other assuming infinite flit buffers at segment switches, are developed. Using approximate Mean Value Analysis, we evaluate performance in terms of processing efficiency and request response time. We also simulate the two models without applying any approximations. We report comparisons of analytical results with simulation results to support the accuracy and appropriateness of our new and novel performance models. The results demonstrate good match between simulation and analytical results and show good scalability for the SMBS. The approach we adopt in developing the models is comprehensive in the sense that the models incorporate features of both direct and nondirect networks. This makes our models easily adaptable to several other network topologies.