Try alternative nucleation algorithm that is consistent across different processor counts.

[fractalsbyx]

Loop over every cell, grab random number, then check if locally owned, rather than looping over locally owned cells. Same seed on every proc.
Slower, but don't know if it is bottleneck.