Application-architecture co-design has emerged as a vehicle for exploring the design space of Exascale architectures. The use of application skeletons as has emerged as a candidate macro-scale structural representation of applications representing the core communication and computation properties of the application. However, they rely on accurate high-level models of the timing behavior of applications to augment the application skeleton. We present a methodology and tools that leverage statistical techniques to allow semi-automated construction of these models and require minimal user insight and architectural expertise. We present results using several Exascale mini-applications that leverage programming frameworks for CPU and GPU architectures.

Slides