The constraints on frequency: 

 - The sampling frequency should be several times the reference frequency 
   to avoid aliasing. I've found that Fs = 8 Fref is satisfactory, and more
   than Fs = 32 Fref makes no difference.

   This is the primary limit on the reference frequency. A low-cost, 
   moderate-data setup -- 200kS/s board, 3 inputs, Fs=32*Fref -- can 
   support reference frequencies up to 2 kHz. With fancy hardware and 
   minimal data requirements -- 1.25 MS/s, 1 channel, Fs=16*Fref -- 
   you can use reference frequencies up to 78 kHz.

   [To make later analysis simpler, I am sampling at a direct multiple of
   the reference.  To go closer to the Nyquist limit of Fs = 2*Fref, use a 
   sampling frequency that is _incommensurate_ with the reference. Otherwise
   the quantization error will start to appear as a separate and nearby 
   signal. Alternately, use a square wave reference, Fs = Fref, and 
   continuous sampling. You will lock in to all the harmonics of the 
   reference frequency, and should adjust the constant of (1/2) in the 
   lock-in procedure.]

 - We scan in a "chunk" of reference waveforms and process the whole batch 
   at once. The samples per chunk should be a power of 2 (1024, 2048, ...) 
   so that FFT's are fast. The faster your PC, the smaller the chunks can 
   be. I've been using 2^14 = 16,384 S/chunk, which is 512 reference cycles 
   for Fs = 32 Fref.

   If this is a problem you can stream to disk for later analysis, or simply 
   do a bare minimum of processing.

 - It's important to avoid discontinuities at the chunk barrier: any 
   discontinuity will appear as noise at the chunk-barrier frequency. 
   If you use either continuous or triggered acquisition there will be no
   discontinuities -- the best solution.  For more work and worse results, 
   you can instead "window" the acquisition and then filter out the 
   windowing frequency.

 - The only lower limit on the reference frequency is the need to sample 
   away from any noise sources. So-called "flicker" noise has a 1/f spectrum:
   it gets progressively worse at lower frequencies.  Frequencies above a 
   couple hundred Hz avoid most of this.  You should also avoid line noise at
   multiples of 60 Hz.  Other than that, try a couple reference frequencies 
   to make sure you don't pick up radio stations or fluorescent light 
   ballasts :-)