Prof. Dr. Torsten Hoefler
Our main goal is to address the three challenges—develop performance bounds, optimization techniques, and applicability to large code-bases—and produce guidelines to enable highest-performance on Intel Xeon Phi™ coprocessors. Furthermore, we pick a specific application, COSMO, in collaboration with domain experts at CSCS and MeteoSwiss. Our optimization methods base on analytic performance modeling which provides most insights for programmers (as opposed to automatic tuning which often acts as a black box).
The project will be carried out in collaboration with the Swiss National Center for Supercomputing (CSCS), which will provide access to computational resources. The second partner, MeteoSwiss, is a member of the COSMO consortium and provides input files and insights into the COSMO code.
The figure below shows an overview of the envisioned tuning technique for the climate and weather code COSMO. The shaded areas (A)-(C) correspond to the research goals: Performance Bounds (A), Application to Large Code Bases (B), Optimization Techniques (C).
- ETH Zurich, 5/1/2017, Capability Models for Manycore Memory Systems: A Case-Study with Xeon Phi KNL, Torsten Hoefler.
- ETH Zurich, 11/19/2014, Programming for Xeon Phi using Cache Line awareness, IXPUG.