The best answer can’t be determined without knowing why 20000 transactions take too long. E.g., it may be because the transactions are applied against very large files with many logicals that must be updated. Breaking the file up might not make much difference if index maintenance is taking the time. A description of the files — physicals and related logicals — is a first step there.
A basic system description (model, memory, disk arms, OS VRM) would be useful too. If memory is a serious problem, running four jobs instead of one might make things worse.
In fact, we don’t even know what “too long” means. Maybe the job takes 30 minutes but the desire is to bring the window down to 15 minutes so a staff member can leave 15 minutes earlier.