User Guide: Low Memory Strategies

Polyorder.jl provides advanced memory optimization strategies enabling massive simulations (e.g., $256^3$ grids) on standard hardware. For detailed benchmarks and algorithm design, see Low Memory Strategies.

User Guide: Low Memory Strategies

Polyorder.jl provides Performance Profiles to easily manage memory optimizations. For large-scale simulations (e.g., $256^3$), simply choosing the right profile is sufficient.

For detailed benchmarks and algorithm design, see Low Memory Strategies.

Quick Start: Selecting a Profile

The profile argument in NoncyclicChainSCFT controls all memory strategies automatically.

Profile	Memory	Speed	Use Case
`:fast`	High	Fastest	Standard CPU jobs ($< 32^3$).
`:balanced`	Medium	Fast	Default. Standard CPU/GPU jobs ($< 128^3$).
`:compact`	Low	Medium	GPU Optimized. Large grids ($128^3 - 256^3$).
`:minimal`	Lowest	Slow	Max Scaling. Extreme grids ($256^3+$) on single GPU.

1. Maximum Compression (Target: $256^3$ Gyroid)

Use :minimal to enable Checkpointing + Symmetric Storage + Shared Caches.

# This fits a 256^3 Gyroid simulation into ~5.6 GB VRAM
scft = NoncyclicChainSCFT(system, w_gpu; profile=:minimal)

2. Balanced GPU Usage (Target: $128^3$)

Use :balanced to enable Symmetric Storage but keep full propagator history (faster than compact and :minimal).

scft = NoncyclicChainSCFT(system, w_gpu; profile=:balanced)

Detailed Features

Feature 1: Time-Domain Checkpointing

(Enabled in :minimal) Reduces propagator storage from $\mathcal{O}(N_s)$ to $\mathcal{O}(\sqrt{N_s})$ by recomputing steps on-the-fly.

Trade-off: ~2x slower simulation.

Feature 2: Space-Domain Symmetry

(Enabled in :balanced, :compact, :minimal) Reduces field storage ($\sim 96\times$ for Gyroid) by exploiting crystal limits.

Note: Polyorder automatically detects symmetry from your UnitCell or AuxiliaryField.
Constraint: Cannot simulate symmetry-breaking transitions (e.g. order-disorder) dynamically.

Troubleshooting

Disabling Symmetry: If you need to study symmetry breaking (e.g., nucleating a new phase), you must disable the automatic symmetry: ```julia
"fast" profile disables symmetry (and precomputes MDEs for speed)
scft = NoncyclicChainSCFT(..., profile=:fast) ```
OOM on GPU?
- Ensure profile=:minimal is used.
- Run GC.gc(); CUDA.reclaim() between solves.
- Use Float32 precision fields.

"fast" profile disables symmetry (and precomputes MDEs for speed)