README_appendix.rst

numericalunits README appendix

Parallelism

Whenever you spawn multiple independent workers solving a common task you have to ensure that the units are the same. You have three possible options here:

• Initialize numericalunits with a common seed.

  # multiprocessing is given for the example
  from numericalunits import reset_units
  from multiprocessing import Pool
  
  reset_units(42)
  pool = multiprocessing.Pool(initializer=reset_units, initargs=(42,))
  

  The above code ensures that all dependent processes in the Pool together with are master process are initialized with the same seed 42. It works only if you have control of the first import of numericalunits.

• A more robust way is to serialize the full content of the numericalunits package.

  import numericalunits
  from multiprocessing import Pool
  
  def serialize_nu():
      return {k: v for k, v in numericalunits.__dict__.items() if isinstance(v, float) and not k.startswith("_")}
  def load_nu(data):
      numericalunits.__dict__.update(data)
  
  # ... some code here ...
  
  state = serialize_nu()
  pool = multiprocessing.Pool(initializer=load_nu, initargs=(state,))
  

• The final option is to ensure a proper conversion. I.e. the data passed to and returned from workers has specific units applied. numericalunits can still be used inside workers.

  from numericalunits import eV, J, kB, K
  from multiprocessing import Pool
  
  energies = eV, 2 * eV, 3 * eV
  
  def worker(energy):
      # Lift units
      energy *= J
  
      # Compute something
      result = energy / kB
  
      # Apply proper units before returning
      return result / K
  
  pool = multiprocessing.Pool()
  temperatures = pool.map(worker, (i / J for i in energies))  # Apply units before passing to worker ...
  temperatures = tuple(i * K for i in temperatures)  # ... and lift units after obtaining the result