Szimuláció és Optimalizáció Szeminárium, 2012. 12. 13.

Gundolf Haase (Karl-Franzens-Universität Graz)

HPC and Mathematics in Applications

2012. 12. 13. 13:30

Location: Járműipari Kutató Központ előadóterem - Inno-Share épület II. emelet

Medium size computer clusters with several hundred CPU-cores are already available to small and medium size companies. The bunch of cores doesn't perform always as well as expected for several reasons and these miracles continue when General Purpose Graphical Processing Units (GPGPUs) are used as additional performance booster. After a brief introduction into recent hardware developments some typical performance bottlenecks caused by physical limitations will be shown. The parallelization of a third party non-Newtonian uid solver on shared memory demonstrates typical pitfalls therein but also very good speedup[1] fi nally.
Especially the GPU parallelization combines all sorts of pitfalls from shared and distributed memory computing. A student project from physics, a random walk for Quantum Chromodynamics (QCD) [2], started with a dissatisfying speedup less than 1 but by taking into account the SIMT (Single Instruction Multiple Threads) hardware of the GPU streaming multiprocessors a much better speedup has been achieved. Further collaborations with students from physics are presented [3]. Some results from an RP-7 project will be presented wherein diff erent goals as GPU-performance and maintainability of an (old) code had to be combined. This problem can be partialy solved by using the pragma-based GPU-compliers that support OpenACC or HMPP.
The situation is di erent for solving sparse systems of equations based on an unstructured discretization of a partial di erential equation. This subproblem has to be solved on a cluster of GPUs in each time step of a cardiovascular simulation. We will show which speedup can be achieved for this sort of problems by using multigrid methods and a proper integration of the parallelization concept into the CARP package1 [4].
We established several cooperations with the local industry during the last two years wherein the HPC competence was the ticket for a start but the mathematics behind is (or will be soon) the real content. These projects don't require always leading-edge-mathematics but mathematics beyond the skills of an engineer.