SRC02. Performance Modeling and Engineering with Kerncraft
Student: Julian Hammer (University of Erlangen-Nuremberg)
Supervisor: Gerhard Wellein (University of Erlangen-Nuremberg)
Abstract: Achieving optimal program performance requires deep insight into the interaction of hardware and software. For software developers without an in-depth background in computer architecture, understanding and fully utilizing modern architectures is close to impossible. Therefore, we must support them with easy to understand models and tools.
We present two simple to apply approaches and a tool for regular stencil and streaming codes: Execution-Cache-Memory (ECM) and layer condition (LC) modeling with Kerncraft. The ECM performance model gives precise overall performance predictions of computational kernels, and the LC analysis gives analytically derived hints for well-suited spatial blocking factors. By utilizing Kerncraft, we are able to take the pain out of the process using automation. We show its applicability to stencils and the insights gained through analysis, performance improvements achieved by following Kerncraft's suggestions derived from generalized LCs, and multi-core saturation point predictions gained through ECM modeling.
Two-page extended abstract: pdf