GTOP: single-threaded CMA-ES through parallel retry
This bundle contains secondary scheduling or target-stopping diagnostics. Every
external cmaes optimizer instance is serial; fcmaes_core::retry supplies only
outer multistart scheduling. See the experiment README and the separate
equal-wall bundle for the user-facing solution-quality comparison.
Fixed-work scheduling diagnostic
The fixed-work phase disables target stopping. Same work counts paired
runs with equal completed retries, evaluations, and final best value. Speedup is
paired sequential wall time divided by parallel wall time. This is a scheduler
diagnostic, not the solution-quality comparison.
| Problem | Workers | Pairs | Same work | Mean speedup | Sdev speedup | Efficiency | Mean active cores | Sdev active cores |
|---|---|---|---|---|---|---|---|---|
| cassini1 | 2 | 1 | 1 | 1.41× | 0.00× | 70.6% | 1.74 | 0.00 |
Fixed-work arm summary
Only the paired external sequential/retry rows above are an exact scheduling comparison. This table exposes the one-start and native-CMA rows without treating their different searches as core-count speedups.
| Problem | Arm | Workers | Runs | Success | Mean best | Sdev best | Mean wall | Sdev wall | Mean evaluations | Sdev evaluations | Mean active cores | Sdev active cores |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| cassini1 | external cmaes, one serial lane | 1 | 1 | 0% | 63.756010 | 0.000000 | 0.0009s | 0.0000s | 207 | 0 | 1.00 | 0.00 |
| cassini1 | external cmaes, sequential retries | 1 | 1 | 0% | 19.719897 | 0.000000 | 0.0035s | 0.0000s | 828 | 0 | 1.00 | 0.00 |
| cassini1 | external cmaes + fcmaes retry | 2 | 1 | 0% | 19.719897 | 0.000000 | 0.0025s | 0.0000s | 828 | 0 | 1.74 | 0.00 |
| cassini1 | fcmaes-core CMA-ES + retry | 2 | 1 | 0% | 26.092117 | 0.000000 | 0.0018s | 0.0000s | 828 | 0 | 1.98 | 0.00 |
Target-oriented results
This phase stops scheduling new retries after reaching the published GTOP target. Already-running starts are allowed to finish, so wall time is the user-visible call duration through worker drain for successes and time to exhaustion for failures. Evaluation counts remain resource accounting, not the primary outcome.
| Problem | Arm | Workers | Runs | Success | Mean best | Sdev best | Mean wall | Sdev wall | Mean evaluations | Sdev evaluations | Mean active cores | Sdev active cores |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| cassini1 | external cmaes, one serial lane | 1 | 1 | 0% | 63.756010 | 0.000000 | 0.0009s | 0.0000s | 207 | 0 | 1.00 | 0.00 |
| cassini1 | external cmaes, sequential retries | 1 | 1 | 0% | 19.719897 | 0.000000 | 0.0034s | 0.0000s | 828 | 0 | 1.00 | 0.00 |
| cassini1 | external cmaes + fcmaes retry | 2 | 1 | 0% | 19.719897 | 0.000000 | 0.0018s | 0.0000s | 828 | 0 | 1.97 | 0.00 |
| cassini1 | fcmaes-core CMA-ES + retry | 2 | 1 | 0% | 26.092117 | 0.000000 | 0.0023s | 0.0000s | 828 | 0 | 1.82 | 0.00 |
Interpretation boundary
The fixed-work comparison only isolates fcmaes_core::retry scheduling.
The fcmaes-core CMA-ES row additionally
changes the optimizer implementation and is not a pure core-count comparison.
The coordinated DE→CMA
results in the parent GTOP report use adaptive budgets and crossover, so they are
a system-level reference rather than another equal-work arm.