I am doing this right now for research into neuro-degenerative diseases, as our supercomputer (we had a BlueGene-P) was decommissioned.
The two most timeconsuming steps as far as I am concerned are:
Breaking the problem down into manageable steps that can be processed in parallel. Obviously if your problem is 'embarrassingly parallel' (that is a technical term guys, i'm not being a douche) this is not a problem.
Wrapping the problem so that you get reasonable performance across many types of hardware. You could also choose to just wrap for GPUs if the sheer number of cores far outweighs the versatility of a CPU. I am still testing this one...