plan.md

Experiment 13: Scatter Access Pattern

1. Focus

• Measure the cost of indirect writes and the slowdown introduced by target collisions.

2. Question

• How much do write-target distribution and contention change scatter throughput?

3. Variants

• unique_targets
• low_collision_random
• high_collision_clustered

4. Method

• Keep the number of logical writes and arithmetic work fixed while changing only the target-index distribution.
• Validate the final output against deterministic CPU reference behavior for each collision regime.

5. Outputs

• Median GPU time by scatter distribution.
• Relative slowdown vs the unique-target baseline.
• Contention-sensitive throughput comparison.

6. Interpretation

• Scatter is not just the write-side mirror of gather because collisions can serialize progress.
• The result should guide whether later pipelines need privatization, staging, or compaction.