Wrapper class for functions which are evaluated by combining (MPI-reducing) a result from each MPI process. See also the distributed_minimize() function.
Example:
import mango.mpi
import mango.optimize
import scipy as sp
import scipy.optimize
def my_func(x):
return (mango.mpi.world.Get_rank()+1) * (x * x + x + 1)
# dfunc sums/reduces (rank+1)*(x*x + x + 1) from all processes to the dfunc.root process
dfunc = mango.optimize.DistributedMetricEvaluator(my_func)
if (dfunc.comm.Get_rank() == dfunc.root):
x0 = 8
res = scipy.optimize.minimize(dfunc, x0, method="Powell")
dfunc.rootTerminate()
print("res.x = %s, res.fun = %s" % (res.x, res.fun))
else:
dfunc.waitForEvaluate()
Methods
| __init__(funcObj[, comm, root]) | |
| calcReduction(localVal) | Converts the local result localVal returned from the evaluate() method to a MPI-reduced result. |
| evaluate(x) | Evaluate the function at x on the local MPI process. |
| rootEvaluate(x) | Broadcasts x to all processes and then does local evaluation (self.evaluate(x)). |
| rootTerminate() | Issues a (self.TERMINATE,x) broadcast so other processes exit the waitForEvaluate() loop. |
| waitForEvaluate() | Loops waiting for an self.EVALUATE broadcast of the x parameter from the self.root-rank process. |
Attributes
| EVALUATE | Instruction to evaluate local MPI function value, see waitForEvaluate() |
| TERMINATE | Instruction to terminate the wait-loop in waitForEvaluate() |