Title:  Analysis of Connection as a Decomposition Technique

Author:  David Daly

Abstract:
Realistic computer systems are hard to model using state-based methods because of the
large state spaces they require and the likely stiffness of the resulting modes (because
activities occur at many time scales).  One way to address this problem is to decompose
a model into submodels, which are solved separately but which exchange results.  We call
modeling formalisms that support the exchange of results between models "connection
formalisms."

This thesis develops connection as a decomposition technique.  The existing connection
infrastructure in the Mobius modeling framework is used to develop a terminology to
describe the decomposition of large models and the inherent associated problems.  A
theory is then developed to describe when such decompositions are feasible.

The theory is then applied to a special class of models to develop a new set of
connection-based approximation techniques that reduce state-space size and solution
time.  This is done by identifying submodels, called isolated submodels that are not
affected by the rest of a model and solving them separatley.  A result from each solved
submodel is then used in the solution of the rest of the model.  We demonstrate the use
of two of these approximation techniques by modeling a real-world file server in the
Mobius modeling framework.  The connected models were solved one to two orders of
magnitude faster than the original model with one of the decomposition techniques
introducing an error of less than 11%.