Parallelism and reproducibility

Many calculations are loops over independent items: k-points for a band structure, energies for a density of states, q-points for a response function, restarts for an annealing. By default they run one after the other. parallel.set_cores(n) runs them on n processes:

from pyqula import parallel
parallel.set_cores(4)          # spread the sweeps over 4 processes

The number of cores changes only the time of the calculation, not the result. This is also true for the calculations that use random numbers, such as the stochastic trace of the kernel polynomial method or the restarts of the classical models: cores=1 and cores=8 give the same numbers. So if you fix the random seed of numpy in your script,

import numpy as np
np.random.seed(42)

the whole calculation gives the same result every time. This is useful when you compare two calculations that differ in one parameter only.

parallel.set_enabled(False) makes everything run on one core, with no parallel processes and no threads. Use it to debug a calculation, or to measure its time on a single core. parallel.set_enabled(True) allows parallel runs again, but it does not restore the previous number of cores, so call parallel.set_cores(n) again.