2 Commits

Author	SHA1	Message	Date
Jukka Aho	c90a028456	perf(benchmarks): Add field storage performance comparison script ... Create 334-line benchmark validating Dict vs type-stable field performance claims from zero_allocation_fields.md design document. Benchmark structure: - Lines 1-18: Header and expected results summary - Lines 20-64: Field type definitions and mock element setup * AbstractField{T}, ConstantField{T}, NodalField{T} * Accessor functions: value(f::ConstantField), value(f::NodalField, node_ids) * Mock element with 8-node connectivity Benchmark suite (5 tests): 1. Constant field access (lines 70-92): Dict["key"] vs value(field) Expected: ~50× faster, 0 allocations 2. Nodal field access (lines 97-120): Array slicing vs @view Expected: ~50× faster, 0 allocations 3. Interpolation without cache (lines 126-170): Type-unstable vs typed Expected: ~16× faster with fewer allocations 4. Interpolation with cache (lines 176-205): Zero-allocation target Uses InterpolationCache struct with pre-allocated result buffer Expected: 0 allocations, maximum speedup 5. Assembly loop (lines 211-261): 1000 elements, Dict vs NamedTuple Expected: 10-100× faster (hoisted constant access) Validation section (lines 267-328): - Compares actual results to claimed performance - ✅/⚠️ status for each benchmark - 10× speedup threshold (conservative vs claimed ~50×) - Zero allocation verification for cached operations Key insights: - Type stability eliminates runtime dispatch overhead - @view and caches achieve zero allocations - Hoisting invariant access provides massive speedup - Validates NamedTuple + typed fields design for v1.0 Dependencies: BenchmarkTools, LinearAlgebra Executable: #!/usr/bin/env julia (chmod +x ready)	2025-11-09 11:07:53 +02:00
Jukka Aho	6a8f8adc1f	docs: Benchmark manual vs AD derivatives for Tet10 ... RESEARCH QUESTION: Should JuliaFEM use hand-calculated derivatives or AD? Created comprehensive benchmark comparing: - Manual: Hand-calculated derivatives (traditional FEM) - AD: Tensors.jl gradient() (automatic differentiation) RESULTS (AMD Ryzen 9, Julia 1.12.1): - Manual: 8.7 ns, 0 allocations - AD: 268.1 ns, 0 allocations - AD is 30× SLOWER than manual KEY FINDINGS: ✅ Both achieve zero allocations (Tensors.jl is well-optimized) ❌ AD has 30× compute overhead from dual number arithmetic ⚠️ In assembly loops: millions of calls = 10+ seconds extra per solve RECOMMENDATION: - Keep manual derivatives for common elements (Tet10, Hex8, Quad4, etc.) - Use AD for prototyping and rare elements - Unit test manual vs AD to catch errors - Future: Generate derivatives symbolically (Symbolics.jl) WHY NOT AD EVERYWHERE? Assembly is hottest path in FEM. 30× overhead = unacceptable for production code. Users will notice the performance difference. WHY NOT ABANDON AD? - Excellent for prototyping - Required for exotic bases (NURBS) - Perfect for unit testing manual derivatives - Zero allocations impressive Files: - benchmarks/tet10_derivatives_benchmark.jl (runnable benchmark) - docs/benchmarks/shape_function_derivatives_ad_vs_manual.md (analysis) Dependencies added: BenchmarkTools This answers the research question definitively with data.	2025-11-09 03:41:47 +02:00