Sniper: Simulation-Based Instruction-Level Statistics for Optimizing Software on Future Architectures

In this paper we address the problem of optimizing applications for future hardware platforms. By using simulation---traditionally a tool for hardware architects---applications, libraries and compilers can be optimized before hardware is available, allowing new machines to start doing useful scientific work more quickly. However, traditional processor simulators are not very user-friendly and are, due to their extreme level of detail, too slow to run applications with large input sets or allow for interactive use. In contrast, the Sniper many-core simulator uses higher abstraction level models, trading off some accuracy for a much higher simulation speed. By adding instrumentation into the simulator that can annotate performance information at fine granularity, down to individual instructions, it becomes a valuable tool for software optimization on future architectures.