Synthetic Program Analysis with Aspen
Jeffrey S. Vetter, Jeremy S. Meredith · 2015
Our community is facing major challenges in the next decade: power, performance, resilience, and productivity. Emerging HPC systems have novel new features, like tightly-integrated heterogeneous computing and nonvolatile memory, that help solve one problem but at the cost of introducing considerable new complexity into the system design. Simply put, we see more complexity and uncertainty in emerging architectures than we have in the last two decades. Given this open design space, architects, applications scientists, and software designers need tools to estimate performance and resource requirements, while taking into account end-to-end design principles. In this paper, we demonstrate how our Aspen performance modeling language can be used to explore important properties of these application design spaces and inform the development of future architectures. We include examples from various applications, and show results from Aspen for idealized concurrency, memory capacity, computational intensity, and others.