Factor isolation and ablation design

Choose factors whose levels correspond to reproducible changes: compression disabled or enabled, two batching limits, or a fixed scheduler policy. Specify how each change is implemented and verify that outputs still satisfy the task contract. Removing a cache may also change memory allocation and synchronization; those are part of that implemented intervention unless separately controlled. A broken ablation that skips required work cannot establish the value of the original mechanism. Include configuration checks so the requested factor level actually reaches the measured path.

For two binary factors A and B, a full factorial includes all four combinations. Code levels as zero and one when interpreting the displayed model; another coding changes coefficient meanings. The coefficient for A describes its effect when B is zero, while the A-by-B coefficient describes how that effect changes when B is one. Do not call the first coefficient an unconditional contribution. For timing outcomes, a negative coefficient can mean a beneficial decrease, so always state the metric and sign convention.

The interaction contrast is a difference of differences. If enabling A saves twenty milliseconds with B absent and fifteen with B present, the savings are conditional rather than additive. A nonzero contrast can reflect resource competition, overlap, or other coupled behavior; the contrast itself does not identify which explanation is correct. Use internal measurements and additional interventions to distinguish those explanations. Report uncertainty for contrasts using the actual run design, including pairing or blocking when applicable.

Randomize run order within suitable blocks to avoid assigning every baseline to a cold morning and every candidate to a hot afternoon. Preserve warmup, reset state, and allocation conditions so preceding treatments do not leak into later cells. With many factors, a fractional factorial can reduce cost, but its alias structure may confound main effects with interactions. Document what the chosen design can estimate separately. Four cell means with no repetition describe the observed contrast but leave no independent within-cell variance estimate for ordinary error analysis.

Pair the factorial table with evidence about the mechanism: bytes transferred, occupied memory, launches, queue lengths, or synchronization. Check that a proposed mediator is measured at the same boundary and does not exclude expensive work moved elsewhere. If two factors cannot be enabled independently, describe that constraint and compare supported configurations instead of inventing a full design. A performance interaction can be scientifically useful even when no simple component attribution exists. Preserve counterexamples where the expected combination does not help.

The roofline instrument offers four exact model cells at a declared main-memory boundary. Reset, set both I=0.5 and B=100, predict 50 GFLOP/s, then Play or Step. Reset, set both I=1 and B=100, predict 100, then Play or Step. Reset, set both I=0.5 and B=200, predict 100, then Play or Step. Reset, set both I=1 and B=200, predict 200, then Play or Step. These are synthetic ceilings min(1000,BI), not measurements or evidence that changing operational intensity is an isolated real-system intervention.