Fields dump load

[kkg4theweb]

First cut of a PR to add support for dumping and (re)loading the 'fields' state.

[codecov-commenter]

Codecov Report

Merging #1674 (80ebafb) into master (89c9349) will decrease coverage by 0.02%.
The diff coverage is 50.00%.

@@            Coverage Diff             @@
##           master    #1674      +/-   ##
==========================================
- Coverage   73.20%   73.17%   -0.03%     
==========================================
  Files          13       13              
  Lines        4515     4519       +4     
==========================================
+ Hits         3305     3307       +2     
- Misses       1210     1212       +2     

Impacted Files	Coverage Δ
python/simulation.py	75.96% <50.00%> (-0.05%)	⬇️

[stevengj]

It's fine to write everything to separate files (one per process, say). You'll want to do a couple of things:

• Modify save_dft_hdf5 to take an additional flag single_parallel_file. If that's false, simply skip the parallel reductions in dft_chunks_Ntotal.
• In your own code, follow the same two-pass structure: first compute the total size of the data on your process, then create the dataset, then write the chunks into that dataset. Again, if single_parallel_file==true we can do an optional parallel reduction after the first pass, and otherwise the code is the same.

Basically, to use HDF5 parallel I/O with a single file you have to create the file and create the dataset collectively, and then each process can call write_chunk as many times as it wants to write its own data into the file (as long as it writes into non-overlapping portions of the file). That's why the save_dft_hdf5 and structure::dump functions are organized the way they are.

[stevengj]

Possible Python API: to re-start a simulation, run the same simulation script, creating the Simulation object as usual with all of the geometry and sources etcetera, but before you do sim.run, first run a sim.load(config) function (that loads the fields, DFT, etcetera).

config would be some kind of argument that says what to load. Maybe a JSON file, or just a dict, whose fields give the various filenames? Alternatively, it could be the name "foo.h5" of a single HDF5 file that contains everything (which maybe in your case is transformed to some other name, e.g. foo__N.h5 for each process N).

[stevengj]

Closed by #1738?