Performance modeling using object-oriented execution-driven simulation

To study the performance of new architectural features, trace-driven simulation is commonly employed. Traditional trace-driven simulation models require the production of an input trace. Depending opt the needs of the simulation study, traces can contain billions of references and consume gigabytes of disk or tape storage. Execution-driven simulation eliminates the need for producing and storing input trace tapes by dynamically generating the input trace stream. The first issue addressed in the paper is the performance of an execution-driven simulation when compared to more traditional trace-driven methodologies. The time and effort spent developing each new model can make a thorough performance evaluation too costly. Instead if a reusable library of architectural simulation models (e.g., caches, branch-predictors etc.) can be constructed using an object-oriented paradigm, the cost of performing the simulation can be drastically reduced. Some of the advantages of using reusable code include: simplified implementation, improved reliability and reduced maintenance. We have developed such a set of execution-driven simulation models which were written using this object-oriented approach. The goal of this work is to provide a reusable framework for rapid prototyping of architecture-based simulation models. The framework is based on the ATOM execution-driven simulation tool provided by DEC. The simulation models are developed using portable C++.