I am trying to do something similar as described in Parallelize a function call with mpi4py But there are some things there that make me skeptical of the provided answer. Additionally, I have a class implementation that makes things a little different. Here is some "reduced" code to demonstrate the main logic: from scipy.optimize import minimize from mpi4py import MPI import numpy as np class Solver(): def __init__(self, SampleTimes, InitialArray): self.comm = MPI.COMM_WORLD self.rank = self.comm.Get_rank() self.size = self.comm.Get_size() self.SampleTimes = SampleTimes self.InitialArray = InitialArray self.Finished = False def f(x_): return <some scalar result> def Objective(self, x): self.x = self.comm.bcast(x if self.rank == 0 else None, root=0) tstep_select = np.array_split(self.SampleTimes, self.size)[self.rank] local_results = [] for t in tstep_select: result = self.f(self.x) # Some function of x local_results.append( (t, result) ) all_results = self.comm.gather(local_results, root=0) if self.rank==0: all_results = [item for sublist in all_results for item in sublist] all_results = np.array(all_results) all_results = all_results[all_results[:,0].argsort()] scalar = np.trapz(all_results[:,1], all_results[:,0]) return -scalar def Maximize(self,): if self.rank == 0: self.Max = minimize(self.Objective, self.InitialArray) self.Finished = self.comm.bcast(True, root=0) return self.Max else: while not self.Finished: self.Objective(None) if __name__=='__main__': t_eval = np.linspace(0, 100, 100) Instance = Solver(SampleTimes=t_eval) print(Instance.Maximize()) So I understand that I need to only call the minimizer on rank 0. Additionally, I need to engage the rank!=0 workers to evaluate Objective. This is the intent of the if/else in the Maximize method. As I understand, I can pass None to Objective for rank!=0 because self.x will be set based on the broadcast from rank 0. I am wondering if the issue has something to do with the fact that scipy's minimize will expect a single value from each evaluation of Objective but, as-written, Objective will be returning None on all rank!=0 (on the other hand, the minimizer is only executed on rank 0 so I feel like this is potentially a non-issue). Continue reading...
