Jakob is a Postdoctoral Research Associate at the Oak Ridge National Laboratory working on performance portable software for fluid dynamics and solid mechanics. He also is a core developer of the Kokkos performance portability library. He got his Ph.D. focusing on numerical fluid mechanics for rotorcraft-ship interaction from the Technical University of Munich in 2024. He also holds a B.Sc. and M.Sc. in mechanical engineering from TUM.
Presentation Title:
Performance Portable Software Design Patterns – Blocks to build the future of high-performance computing software
Presentation Abstract:
The future of science will continue to rely on software to compute predictions, plan studies, evaluate results, and validate the already known. But scientific software is exceedingly complex. At the same time, it needs to be correct to allow scientists to draw conclusions.
To manage the complexity and correctness of software, humans often make complexity local, explicit, and testable. This strategy continues to be valuable in the age of AI assisted or AI lead development of scientific software.
Software design patterns have been very successful at encapsulating complexity, being explicit about the intent, and highlighting what needs to be tested to give correctness guarantees.
But many design patterns are still CPU focused and do not work on GPUs. To solve this, the presented effort focuses on abstracting performance portable design patterns from existing code developed at ORNL/DoE.
It will highlight various patterns that have been generalized and outline how they can help software development in the age of AI.
This work is done as part of a BSSw Fellowship.