TDM data has a default integration time of 2 seconds, FDM data has 6 seconds. That means that you quickly build up a lot of data, which can be significantly reduced by suitable pre-averaging. Usually, binning in time is a good idea (unless you want to have self-calibration solutions on very short timescales). This can be done by splitting out data, using the timebin keyword:

split(vis = 'ISON.ms',
  outputvis = 'ISON_timebin.ms',
  datacolumn = 'data',
  timebin = '30s')

CASA has a very powerful continuum imaging mode, called Multifrequency Synthesis. This takes continuum data from all your spectral windows at different frequencies and generates a continuum image at a single frequency. However, this mode can be quite slow if you have a lot of channels. Also here, pre-averaging is a good idea.

Typically, around 10 channels per spw are more than enough. So a FDM window with 3840 channels can easily be averaged by a factor of a few hundred. In the following example, we have a measurement set with 3 FDM windows (3840 channels each) and a TDM window (128 channels). We will average them all to have 8 channels per spw:

split(vis = 'ISON.ms',
  outputvis = 'ISON.ms.continuum',
  datacolumn = 'data',
  width = [480, 480, 480, 16])