Data Oriented Design Guide

Data-Oriented Design Guidelines

Essentials

• The problem is data movement - Optimize cache traffic, not instruction count, see references/cache-behavior.md
• Design from the data - Model bulk input→output transforms, not idealized objects, see references/data-as-transforms.md
• Measure first - Profile cache misses before changing layout, see references/measurement-and-profiling.md
• Record cheaply, always on - Per-frame accumulating counters recorded via a cached accumulator pointer (one store, no lookup/branch/lock), history allocated only for viewed counters, see references/statistics-recording.md

Layout

• SoA / AoS / AoSoA - Split structs into per-field arrays for hot loops, see references/soa-aos-aosoa.md
• SIMD-friendly layout - Contiguous columns, padding, alignment for vectorization, see references/simd-friendly-layout.md
• Nested / variable-length arrays - Hold child lists in bulk arrays + cache-line chunks, not per-object heap pointers, see references/nested-arrays.md

Processing

• Sequential access - Stream linearly, never chase pointers, see references/access-patterns.md
• Existence-based processing - Bucket by state so branches become loops, see references/existence-based-processing.md
• Batch over one-at-a-time - Transform N items per call, amortize overhead, see references/data-as-transforms.md
• Default to zero - Make all-zero a valid default; reserve 0 for none/neutral, not magic sentinels, see references/zero-as-default.md

Memory

• Handles, not pointers - Reference by index/handle into relocatable, stable arrays (swap-remove, free lists, indirection tables), see references/handles-and-indices.md
• Allocation - Contiguous storage matters for cache, but the allocators themselves (arenas/pools/lifetimes) are general, see memory-management-guide

Gotchas

• SoA only wins when loops touch a subset of fields; full-record access can favor AoS: measure both.
• The hardware prefetcher tracks linear strides; randomizing your index order silently disables it.
• Padding for alignment trades memory for throughput, on cache-bound loads, the smaller packed layout can still win.
• A counter you cache as a raw pointer dangles if its backing array reallocates: hand out indices into a stable block, or pointers into a non-relocating pool.

Progressive Disclosure

• Read references/cache-behavior.md - Load when reasoning about cache lines, latency, or miss types
• Read references/data-as-transforms.md - Load when modeling a system as bulk streams instead of objects
• Read references/soa-aos-aosoa.md - Load when choosing or converting between AoS, SoA, and AoSoA layouts
• Read references/access-patterns.md - Load when iterating collections or replacing pointer-chasing structures
• Read references/handles-and-indices.md - Load when designing entity references, stable arrays, or free lists
• Read references/existence-based-processing.md - Load when removing per-item branches by sorting or bucketing
• Read references/simd-friendly-layout.md - Load when laying out data for vectorization or alignment
• Read references/nested-arrays.md - Load when objects own variable-length child lists, or you're reaching for a per-object growable container
• Read references/zero-as-default.md - Load when choosing sentinels, nil values, or default-initialized state
• Read references/measurement-and-profiling.md - Load when measuring before/after or reading hardware counters
• Read references/statistics-recording.md - Load when building always-on counters/telemetry that must add near-zero overhead on hot paths