I was just realizing something along these lines. It would be better to give a larger proportion of disks to the subpiles which will have more scratch space. How much more, I don't exactly know (yet). I've been trying to come up with a simple way to minimize the recurrence for the general case, but no luck.
If all else fails, you could search through all partitions of N-1 to find the best subproblem split, but that's no fun ;)