Potential Performance Improvements for LST binning

[steven-murray]

This issue is just to keep a little log of performance for the lst binner as it stands, and to serve as a jumping-off point for future improvements, if necessary.

When run on redundantly-averaged H6C data (1760 baseline-pols), the lst_bin_files function takes 337 seconds (5.5 min) per output file (with 2 10-second LST bins per file).

The major breakdown of time is:

• 1min [~20%] -- reading the LST bin configuration YAML
• 64sec [~20%] -- reduce_lst_bins (i.e. doing the averaging itself)
• 51sec [~18%] -- HERAData.read() (no, not using the Fast reader yet)
• 2 min [~40%] -- keyed_on_bls() (i.e. finding the unique baselines that exist in the data)

Of these, I think the LST bin configuration YAML reading and the unique baseline keying should be simplest to trim down. I have no reason to suspect that they shouldn't be roughly negligible, meaning we can ~halve the time of LST binning.

For NON redundantly averaged H6C data, the lst_bin_files function takes 5150 seconds (~1.5 hours) per output file.
The major breakdown of time is:

• 64% in lst_bin_files_for_baselines, of which
  • 73% in reading data (HERAData.read), of which
    • 50% in actual uvdata read
    • 20% in blt slice determination
    • 30% in build_datacontainers
  • 10% calibrate-in-place
  • 6% in lst-align
• 32% in reduce_lst_bins , of which
  • 37% in lst_average
  • 60% in getting MED/MAD (which was supposed to be turned off...)