the export is the fastest if you run export with direct=y recordlength=65000
but import will possibly last 45 min for every GB. And then it is best
to import with indexes=n
and make the redologs as large as you can to avoid many log switches.
after the job has finished loading the rows and primary indexes you can submit a job to build the indexes in parallel and no logging - therefore it necessary to manipulate the generated file for the indexes with an sed command to replace the string logging with nologging parallel
> the export is the fastest if you run export with direct=y
Just so you know: I have done exports on an Oracle 7.3.4/SCO-Unix V5 system and these exports were dramatically slower with the "direct=y" option. With "direct=n" speed was about 15 min./GB. I did not pursue this any further at the time.