 Jukka Aho
|
81f4f85f3e
|
feat(gpu): Multi-GPU MPI benchmark with nodal assembly
Implements working GPU-accelerated nodal assembly with MPI domain decomposition:
- Matrix-free matvec operation on GPU (y = A*x without assembling A)
- 2-6× speedup vs CPU multi-threading (114-302 Mnodes/s)
- Scales to 343K nodes / 1M DOFs with acceptable communication overhead
- CSR format for GPU-friendly node-to-elements connectivity
- Global-to-local index remapping for partition consistency
Key components:
- benchmarks/multigpu_mpi_benchmark.jl: Full MPI+CUDA implementation (555 lines)
- benchmarks/multigpu_results_2025-11-09.md: Detailed performance analysis
- docs/book/gpu_benchmark_milestone.md: Comprehensive tutorial documentation
Performance results (NVIDIA RTX A2000 12GB, 2 MPI ranks):
- 30³ mesh: 114.84 Mnodes/s, 29% communication overhead
- 50³ mesh: 130.64 Mnodes/s, 61% communication overhead
- 70³ mesh: 301.83 Mnodes/s, 51% communication overhead
Architecture validated: Nodal assembly + matrix-free + GPU = fast and scalable.
Foundation complete for production FEM solver (needs: real stiffness, GMRES, preconditioner).
|
2025-11-09 15:59:00 +02:00 |
|