Jukka Aho
69e3b131b5
GPU port of nodal assembly strategy with CUDA kernels demonstrating atomic-free assembly on GPU using node-parallel approach. GPU kernel design: - One thread per node (not per element) - Each thread gathers from touching elements - No atomic operations (node ownership) - Coalesced memory access via node ordering Kernel structure: - Thread ID maps to node ID - Loop over elements touching this node - Loop over element nodes for block contributions - Compute 3×3 stiffness blocks with Tensors.jl - Accumulate locally, write once to global Data layout: - node_to_elements: CSR-like structure on GPU - Element data: Array of Structs (immutable elements) - Node displacement: Flat vector (3*n_nodes) - Result: Flat vector (3*n_nodes) Performance characteristics: - Memory bandwidth bound (not compute bound) - Benefits from coalescing (sequential node access) - Scalable to multi-GPU (domain decomposition) - No synchronization within kernel Comparison to element assembly: - Element: N_elem threads, atomic scatter - Nodal: N_nodes threads, no atomics Reference: CPU version in nodal_assembly_cpu.jl
JuliaFEM Technology Demonstrations
This directory contains demonstrations of key technologies and architectural decisions for JuliaFEM v1.0.
Overview
These demos validate that type-stable field storage enables modern high-performance computing patterns: GPU execution, MPI communication, and Krylov iterative solvers.
Demonstrations
1. GPU and MPI Communication (gpu_mpi_demo.jl)
Purpose: Prove that type-stable data structures flow efficiently to GPU and MPI.
What it demonstrates:
- Real CUDA GPU kernel execution
- MPI data transfer between processes
- Combined GPU+MPI workflow
Run:
mpiexec -np 2 julia --project=. demos/gpu_mpi_demo.jl
Documentation: README_GPU_MPI.md
2. Multi-GPU MPI Krylov Solver (krylov_mpi_gpu_demo.jl)
Purpose: Complete distributed FEM solver workflow with nodal assembly.
What it demonstrates:
- Nodal assembly pattern (row-by-row matrix construction)
- Distributed matrix-vector products
- Conjugate Gradient solver with MPI
- Multi-GPU execution
- Solution verification (10×10 SPD system)
Run:
mpiexec -np 2 julia --project=. demos/krylov_mpi_gpu_demo.jl
Documentation: README_KRYLOV_DEMO.md
Results:
- Converges in 9 iterations
- Relative error: 7.73 × 10⁻¹⁴
- Validates complete distributed solving workflow
Requirements
Required
- Julia 1.9+
- MPI installation (e.g., OpenMPI, MPICH)
- MPI.jl package
Optional (for GPU demos)
- CUDA-capable GPU
- CUDA.jl package
If CUDA is not available, demos will fall back to CPU execution while still demonstrating the distributed computing patterns.
Key Insights
Type Stability is Not Optional
These demos prove that type-stable field storage is required (not just "nice to have") for:
| Feature | Why Type Stability Required |
|---|---|
| GPU execution | CUDA kernels cannot compile with abstract types |
| Fast MPI | Typed buffers avoid serialization overhead |
| Krylov solvers | Matrix-free operators need concrete types |
| CPU performance | 9-92× speedup measured (see CPU benchmarks) |
Nodal Assembly Pattern
The Krylov demo shows row-by-row matrix construction (get_row() abstraction), which:
- Aligns naturally with contact mechanics (nodal constraints)
- Enables domain decomposition (each rank owns nodes)
- Supports matrix-free solving (never form global matrix)
- Scales to large problems (O(N) memory vs O(N²))
Architectural Validation
These demos validate the design decisions for JuliaFEM v1.0:
v0.5.1 (2019):
- Dict-based fields → Type instability
- Element assembly → Global matrix
- Direct solvers → O(N³) time, O(N²) memory
- Single-threaded CPU
v1.0 (target, validated here):
- Type-stable fields → GPU/MPI capable
- Nodal assembly → Distributed construction
- Krylov solvers → O(N·iter) time, O(N) memory
- Multi-GPU + MPI
References
- CPU benchmarks: See
benchmarks/field_storage_comparison.jlfor 9-92× speedup measurements - Design documentation: See
docs/book/zero_allocation_fields_v2.mdfor architectural rationale - Session notes: See
llm/sessions/2025-11-09_gpu_mpi_validation.mdfor development history
Contributing
These demos are educational and meant to be:
- Clear: Understand what's being demonstrated
- Minimal: No unnecessary complexity
- Runnable: Work on typical hardware (fall back to CPU if needed)
- Validated: Compare against exact solutions
When adding new demos, follow this pattern and document thoroughly.