Browse Theses

The use of trace-driven simulation in the evaluation of computer memory systems has been popular for at least 30 years. Despite considerable progress in the development of this method, the best trace-driven simulators still run at least one to two orders of magnitude slower than actual hardware and it is unlikely that they will exhibit further substantial improvements in speed.

Trap-driven simulation is a new method for memory-system simulation that overcomes the bottlenecks inherent in trace-driven simulation speed. By invoking a memory-system simulator only on references that cause a change in the simulated memory state, trap-driven simulators can potentially reduce memory-system slowdowns to zero. Although promising in principle, little is known about how trap-driven simulation compares, in practice, to trace-driven simulation.

Through a literature survey of trace-driven simulation and the implementation of a prototype trap-driven simulator, Tapeworm II, this dissertation compares the trace-driven and trap-driven methods on the basis of their flexibility, portability, speed, and accuracy. We show that trap-driven simulation offers clear advantages over trace-driven simulation with respect to speed and accuracy. With some exceptions, trap-driven simulation is nearly as flexible as trace-driven simulation, but suffers from certain problems of portability. We suggest methods for alleviating these problems, through simulation-host hardware support in existing and future systems.