2015 Fiscal Year Annual Research Report
ＧＰＵによって加速化された氷床モデリング用プログラムの開発と応用
| Project/Area Number |
15F15902
|
|---|
| Research Institution | Hokkaido University |
|---|
Principal Investigator |
Greve Ralf 北海道大学, 低温科学研究所, 教授 (90374644)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
SEDDIK HAKIME 北海道大学, 低温科学研究所, 外国人特別研究員
|
|---|
| Project Period (FY) |
2015-04-24 – 2017-03-31
|
| Keywords | Ice sheet / Dynamics / Full Stokes / Numerical model / GPU computing / Finite element method / Matrix assembly |
|---|
| Outline of Annual Research Achievements |
Our research aims at developing an experimental ice flow model which employs GPUs (graphics processing units) to accelerate the computationally expensive finite element method. During the first year of the project, the work focused on implementing the full-Stokes solver for both the CPU (serial mode) and the GPU (parallel mode). The CPU implementation is generic and targets multiple flow problems (e.g., compressible or incompressible flow), while the GPU implementation only targets ice flow modeling in ice sheets. The GPU implementation was devised so that data transfers to the device and branching code are kept low in order to maximize performance. A computing scheme for the global matrix assembly on the GPU was implemented using a coloring algorithm which groups all mesh elements with no sharing nodes into the same colored bucket.
The new model was implemented in modern object-oriented programming and was enhanced to solve coupled problems for both steady-state and transient conditions. A mesh extrusion procedure and full support for periodic boundary conditions were also implemented. In order to verify the correctness and robustness of the model, we tested the full-Stokes solver using the Ice Sheet Model Intercomparison Project ISMIP-HOM. Two experiments were considered: ice flow over a bumpy bed (Experiment A) and ice flow over a rippled bed (Experiment B). For both experiments, the CPU implementation of the full-Stokes solver gave correct and reproducible results.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
We originally planned to have in the first year both the CPU- and GPU-based full-Stokes solvers implemented and tested with the ISMIP-HOM benchmarks. We achieved this goal so far only for the CPU-based implementation. The GPU-based solver was fully implemented but not tested yet with the benchmarks. The delay is due to more time spent in reworking the code base so that it matches better the most recent versions of Elmer, which is the open-source multi-physics finite element software from which our new model was derived. This rework was also necessary to easily use the Elmer input files written for the ISMIP-HOM tests. Additionally, the time required to properly implement periodicity in boundary conditions was longer than originally expected.
|
| Strategy for Future Research Activity |
During the second year of the project, we plan to fully test the GPU-based full-Stokes solver. For this propose, we will use the same ISMIP-HOM experiments as for the CPU code. The tests will be used to assess the correctness of the GPU implementation and to devise optimization techniques for the GPU code for maximum efficiency. In particular, it will be crucial to carry out the experiments in order to assess the most effective work-group size of computing kernels for optimum utilization of the GPU local memory. Additionally, the CPU and GPU implementations of the full-Stokes solver will be enhanced to successfully compute two more ISMIP-HOM tests: ice stream flow I (Experiment C) and ice stream flow II (Experiment D). For this purpose, a few minor additions should be made to the code base in order to properly take into account basal sliding conditions.
The optimized GPU code will be compared to the model Elmer/Ice using again ISMIP-HOM. In order to stress both models, a high-resolution finite element mesh will be used for each experiment. Elmer/Ice is written in Fortran and parallelized using Open MPI. For that reason, it is a fast and proven reference implementation to which our new GPU-based model can be compared in order to assess the efficiency of GPUs for the purpose of ice sheet modeling. We finally hope to share our findings in a peer-reviewed journal.
|
Research Products
(1 results)
