Application - Architecture Interaction on Shared

Programmers who need to obtain high performance on their shared-memory multiprocessor applications are already required to modify and sometimes rewrite applications. Running low-level simulations allows the programmer to more easily see performance bottlenecks and to measure more characteristics of an application running on an architecture, including characteristics that are impractical for hardware to measure. Low-level simulation is more accurate than high-level simulation and helps prevent unrealistic assumptions from being made and allows characterizations to be made that cannot be made by existing multiprocessor hardware. We describe the Multi-multi shared-memory multiprocessor architecture and show several examples where we determined the bottleneck of an application running on the architecture by using low-level simulation. Performance bottlenecks discovered in simulation of an application on a speciic shared-memory multiprocessor are often generalizable to achieve high performance of the application on a wide variety of shared-memory multiprocessors.