[Discussion] Performance of cached vs non-cached geometry

[Yurlungur]

I went ahead and ran the torus for 100 cycles with and without cached geometry to compare performance. Cache wins... barely. Some details:

Setup

• I used a single 256*256*256 meshblock.
• For the cached geometry, I turned on axisymmetric.
• I ran on a single A100 GPU.
• I ran for 100 cycles.
• Otherwise I used the default torus.pin input deck.

Results

• The cached geometry ran at 5.75e7 zone-cycles/gpu-second
• The analytic geometry ran at 4.0e7 zone-cycles/gpu-second

This translates to about a 30% speedup between cached and analytic geometries in 3D... with the cached geometry being faster.