If at all possible, do prepare all files needed for your job on your machine. When everything looks OK, ftp them across and start your job. You wouldn't want to see root's icy eyes up-close-and-personal because a parallel 32-core job can't start just because you have a series of 5-minute serial jobs stacked-up in the queue. There is an important exception to this rule: you should tune and calibrate your calculations for Norma. The quantitative way to do this is to calculate the job's parallel efficiency as described
here. For large and expensive clusters the rule of thumb is to never go below 70% efficiency, but we (the utmost periphery universities) can bend that a bit