plan.md

Experiment 14: Read Reuse and Cache Locality

1. Focus

• Measure how temporal reuse distance changes read performance on the same logical data.

2. Question

• How quickly does the benefit of a second read decay as more work is inserted between touches?

3. Variants

• reuse_distance_1
• reuse_distance_32
• reuse_distance_256
• reuse_distance_4096
• reuse_distance_full_span

4. Method

• Build deterministic index schedules where each source element is read twice with a controlled gap.
• Keep total reads, writes, and arithmetic fixed while varying only the reuse distance.

5. Outputs

• Median GPU time by reuse distance.
• Speedup relative to the far-distance baseline.
• Stability of the locality benefit across the reuse sweep.

6. Interpretation

• This experiment estimates a practical locality window rather than a literal cache size.
• It helps explain when reordering work or staging data is likely to matter.