SciSim: a software performance estimation framework using source code instrumentation

Recently, software performance estimation based on source code instrumentation shows promising results in the literature. It achieves significant speedup without compromising accuracy, compared with cycle-accurate simulations. However, much work still remains to be done to make this technique flexible and accurate enough to estimate software on complex processors. To the best of our knowledge, we are the first to propose ways to tackle microarchitecture related issues in the source code instrumentation approach. We perform static instruction scheduling for superscalar architectures at instrumentation time and combine instrumented code and microarchitecture simulators to model runtime interactions between software and microarchitecture. We have developed a new framework, SciSim, to provide a common infrastructure for the proposed approach. It is designed to be easily extendable and retargetable to different instruction set architectures and processors. Using SciSim SystemC modules may be automatically generated to integrate software into system-level simulation. We will present the applicability of SciSim in system-level design exploration of multiprocessor systems. At last, experiments with standard benchmarks are presented to validate the speed and accuracy of SciSim.