We find that the scanline strategy is usually optimal when individual read operations are cheap (i.e. the source data is held locally in an uncompressed format) or where source grids are very large (for which the bounding box strategy would require very large amounts of memory).
The bounding-box algorithm is most useful where the data are held in a compressed format (such as HDF5 or NetCDF4), or where the data are held on a remote OPeNDAP server; in these cases, individual read operations carry a large overhead.