Analysis of simulation-adapted SPEC 2000 benchmarks

In this paper we have compared the accuracy and representativeness of two common approaches (namely, fast forwarding and reduced input datasets) employed in current architectural studies to deal with the extremely long times involved in detailed architectural simulations. The experimental methodology chosen to perform the validation of both techniques consists in direct execution and performance monitoring through hardware performance counters, our study being focused on the nine SPECint benchmarks for which a reduced input dataset version exists. As a target system we have opted to use an Intel Pentium-4 microprocessor due to its remarkable monitoring features. The results obtained suggest that both techniques are similar, i.e. in some cases Fast-forward is preferable and vice versa, whereas for some benchmarks neither of the techniques are able to provide satisfactory results. However, the simulation cost associated with the fast-forwarding technique is about two orders of magnitude lower than that corresponding to reduced input datasets.