I'm not quite sure who build the machine for you originally. You might try to contact them and ask them about the tuning. The tuning is very much motor/machine/encoder depended. There is not a universal procedure.
When you going to tune the R990H (R990H/R991H/R992H/R90H) drive you will have to run program such as:
http://www.rutex.com/zip/RT990H.zip
or latest version of Mach2
The older R990 (R90/R990/R995/R996/R997) drive can be tuned with the:
http://www.rutex.com/zip/R90.EXE
The reason for that is that the older version do the calculations in 16-bit and the 'H drives in 32-bit.
Before you start to tune drive you should mechanically disconnect motor from the machine and you should have the E-stop switch very close to your hand. Incorrectly tuned drive can violently oscillate which can do mechanical damage to the machine/motor or overheat the drive. As well, it is good idea to start the tuning at lower voltage (if you can drop down the voltage) to families yourself with tuning. There is much less mechanical/electrical energy.
Secondly, you should know if your system needs the Ki or not.
The difference is that without Ki the system is bit faster - sort of 'snappier' when traveling, but there is a inherent error of few encoder counts - somewhere around 5-50 encoder counts. With Ki the system might be slightly less dynamic but it will correct the steady error down to +/-1 encoder count. Most of the application is tuned with Ki.
Typical setup for DC PM running at DC30V with 500-line encoder could be somewhere Kp=1000, Ki=0, Kd=2000. If you apply Ki, then the Kp and Kd has to be dramatically increased. As well, the Ki has to be adjusted over certain value to work correctly. Usually very low Ki does no do good job at all. So the setting could be somewhere Kp=5000, Ki=150, Kd=3000. If you double the Voltage and double the encoder resolution, then the Kx values should be about the half. The above values are very bulk figures and you might end up with completely different.
When you look at the step response graph output, then it should be tuned for fastest setting time to acceptable error and not the zero overshoot. The step response does not have preprogrammed acceleration/deceleration so it can or it should overshoot.
Cheers
vh