Reference sources should generally be chosen from the VLA Calibrator List (maintained by the NRAO staff and referenced by the OBSERVE program), unless you are sure that a putative calibrator is unresolved in the VLA configuration that will be used, and has a position measured in the VLA reference system to better than 0.1''.
The length of time spent on each calibration scan should be enough to achieve a signal-to-noise (over the 26 baselines contributing to each antenna gain solution) commensurate with the required calibration accuracy. Never plan to calibrate for less than 1 minute at a time, however, as shorter calibrator scans may be lost as a result of unusually long settle-down times, etc. Significant atmospheric amplitude and phase fluctuations can occur on time scales of minutes, even at wavelengths of 6 cm and longer. When the Sun is active, ionospheric fluctuations will dominate at 18 cm and longer--they can also be rapid on long VLA baselines. It is completely impractical to adopt a calibrator-target-calibrator cycle that will guarantee following the fastest fluctuations of either kind. Keep in mind that no external referencing, no matter how rapid, can be guaranteed to remove atmospheric fluctuations from the data, and that time spent driving to and observing calibrators is time deleted from integration, and u-v coverage, on your target. You must decide for yourself how to play this particular roulette game during a given run.
Typical VLA observing programs spend from 5% to 10% of their time on calibration at the lower frequencies; more calibration may be needed at the higher frequencies where the calibrators are weaker and therefore need to be observed for longer total integration times. Calibration every 20 minutes or so will often follow the longer-term atmospheric fluctuations at 20 cm and 6 cm, especially in the more compact VLA configurations. Calibration every 10 minutes or so is safer at 2 cm and shorter wavelengths, especially if the external calibrator is within a degree or so of the target. It should not be necessary to calibrate instrumental effects more rapidly than every 30 minutes at 20 cm or 6 cm.
Finally, note the significance of the choice of the gain table
interval for the VLA off-line data base created by the FILLM
program if you will not self-calibrate your data. The off-line
gain table interval (which you specify to the array operator at the
time of the observations) sets the minimum time scale of instrumental
or atmospheric fluctuations that can be corrected by an external
calibration. (Self-calibration algorithms construct their own gain
tables based on the integration time 
that you specify for the gain determination.) The VLA default gain
table interval of 10 minutes is adequate for calibration under a
stable atmosphere, but shorter intervals are often more appropriate if
you will rely on external calibration.