From what you say, I understand that you want to decrease the df of your spectrum. You know that df = 1/(N dt), where N is the number of your samples and dt = 1/fs is the sampling interval. Therefore, df = 1/T where T = N dt is the total time of observation.
So in order to make df finer, you need to increase the time of recording, T.
When you just zero-pad (as when you take the 2^nextpow2), your figure will have a finer df but in fact the resolution hasn't increased. fft just "interpolates" the values.