I´ve got funny problem and no explanation/solution:

I run a JCL with three steps: sort a file, unload a table, run program. run time is really high (just a few seconds CPU time and more than half an hour).

If I divide the three steps into three single jcl and submit jcl, every jcl just need a few seconds. So running three jcl is much more faster than one jcl with three steps.

what can be a reason? Do I have to set some special parameters for scheduling, to get good run time in both cases?

Any single JCL job runs its' programs sequentially. If you're familiar with the lower platforms, it's like running a script. No line can run until the previous line has completed.
Only in this case, no program can run until the previous one has completed.

When you submit 3 jobs, each gets a different initiator and run simultaneously. In fact, the ability to run many, many jobs simultaneously is the hallmark of a mainframe.
On a lower platform, this is similar to forking several processes.
I think if you add up all the time from the 3 separate jobs, you will have slightly higher runtimes and resource usage (wall clock, SRB, EXCP & CPU) than with a single job.

I'm not sure I understand your question regarding scheduling. If the first step creates a dataset that the second step uses and the second then creates a dataset used in the third, you have no other recourse than to single-thread these jobs.
If the three steps have nothing to do with one another, then submitting them separately is a good idea.

In theory, you could use BDT and batch pipes, but based on your question, I’d say you need to gain some more experience before working with those.

kk