"""
Baseclass for design-of-experiments Drivers that have pre-determined
parameter sets.
"""
import os
import traceback
from six.moves import zip
from six import next
import numpy
from openmdao.core.driver import Driver
from openmdao.util.record_util import create_local_meta, update_local_meta
from openmdao.util.array_util import evenly_distrib_idxs
from openmdao.core.mpi_wrap import MPI, debug
trace = os.environ.get('OPENMDAO_TRACE')
[docs]class PredeterminedRunsDriver(Driver):
"""
Baseclass for design-of-experiments Drivers that have pre-determined
parameter sets.
Args
----
num_par_doe : int, optional
The number of DOE cases to run concurrently. Defaults to 1.
load_balance : bool, Optional
If True, use rank 0 as master and load balance cases among all of the
other ranks. Default is False.
"""
def __init__(self, num_par_doe=1, load_balance=False):
if type(self) == PredeterminedRunsDriver:
raise Exception('PredeterminedRunsDriver is an abstract class')
super(PredeterminedRunsDriver, self).__init__()
self._num_par_doe = num_par_doe
self._par_doe_id = 0
self._load_balance = load_balance
def _setup_communicators(self, comm, parent_dir):
"""
Assign a communicator to the root `System`.
Args
----
comm : an MPI communicator (real or fake)
The communicator being offered by the Problem.
parent_dir : str
Absolute dir of parent `System`.
"""
root = self.root
if self._num_par_doe < 1:
raise ValueError("'%s': _num_par_doe must be >= 1 but value is %s." %
(self.pathname, self._num_par_doe))
if not MPI:
self._num_par_doe = 1
self._full_comm = comm
# figure out which parallel DOE we are associated with
if self._num_par_doe > 1:
minprocs, maxprocs = root.get_req_procs()
if self._load_balance:
sizes, offsets = evenly_distrib_idxs(self._num_par_doe-1,
comm.size-1)
sizes = [1]+list(sizes)
offsets = [0]+[o+1 for o in offsets]
else:
sizes, offsets = evenly_distrib_idxs(self._num_par_doe,
comm.size)
# a 'color' is assigned to each subsystem, with
# an entry for each processor it will be given
# e.g. [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3]
color = []
self._id_map = {}
for i in range(self._num_par_doe):
color.extend([i]*sizes[i])
self._id_map[i] = (sizes[i], offsets[i])
self._par_doe_id = color[comm.rank]
if self._load_balance:
self._casecomm = None
else:
casecolor = []
for i in range(self._num_par_doe):
if sizes[i] > 0:
casecolor.append(1)
casecolor.extend([MPI.UNDEFINED]*(sizes[i]-1))
# we need a comm that has all the 0 ranks of the subcomms so
# we can gather multiple cases run as part of parallel DOE.
if trace:
debug('%s: splitting casecomm, doe_id=%s' % ('.'.join((root.pathname,
'driver')),
self._par_doe_id))
self._casecomm = comm.Split(casecolor[comm.rank])
if trace: debug('%s: casecomm split done' % '.'.join((root.pathname,
'driver')))
if self._casecomm == MPI.COMM_NULL:
self._casecomm = None
# create a sub-communicator for each color and
# get the one assigned to our color/process
if trace:
debug('%s: splitting comm, doe_id=%s' % ('.'.join((root.pathname,
'driver')),
self._par_doe_id))
comm = comm.Split(self._par_doe_id)
if trace: debug('%s: comm split done' % '.'.join((root.pathname,
'driver')))
else:
self._casecomm = None
# tell RecordingManager it needs to do a multicase gather
self.recorders._casecomm = self._casecomm
root._setup_communicators(comm, parent_dir)
[docs] def get_req_procs(self):
"""
Returns
-------
tuple
A tuple of the form (min_procs, max_procs), indicating the
min and max processors usable by this `Driver`.
"""
minprocs, maxprocs = self.root.get_req_procs()
minprocs *= self._num_par_doe
if maxprocs is not None:
maxprocs *= self._num_par_doe
return (minprocs, maxprocs)
[docs] def run(self, problem):
"""Build a runlist and execute the Problem for each set of generated
parameters.
"""
self.iter_count = 0
if MPI and self._num_par_doe > 1:
if self._load_balance:
runlist = self._distrib_lb_build_runlist()
else:
runlist = self._get_case_w_nones(self._distrib_build_runlist())
else:
runlist = self._build_runlist()
with problem.root._dircontext:
# For each runlist entry, run the system and record the results
for case in runlist:
if MPI and self._load_balance and self._full_comm.rank == 0:
# we're the master rank and case is a completed case
self.recorders.record_case(problem.root, case)
elif case is not None: # dummy cases have case == None
metadata = create_local_meta(None, 'Driver')
update_local_meta(metadata, (self.iter_count,))
for dv_name, dv_val in case:
self.set_desvar(dv_name, dv_val)
problem.root.solve_nonlinear(metadata=metadata)
if self._load_balance:
# keep meta for worker to send to master
self._last_meta = metadata
if not MPI or not self._load_balance:
self.recorders.record_iteration(problem.root, metadata,
dummy=(case is None))
self.iter_count += 1
def _get_case_w_nones(self, it):
"""A wrapper around a case generator that returns None cases if
any of the other members of the MPI comm have any cases left to run,
so that we can prevent hanging calls to gather.
"""
comm = self._casecomm
if comm is None:
for case in it:
yield case
else:
cases_remain = numpy.array(1, dtype=int)
while True:
try:
case = next(it)
except StopIteration:
case = None
val = 1 if case is not None else 0
comm.Allreduce(numpy.array(val, dtype=int),
cases_remain, op=MPI.SUM)
if cases_remain > 0:
yield case
else:
break
def _distrib_build_runlist(self):
"""
Returns an iterator over only those cases meant to execute
in the current rank as part of a parallel DOE. _build_runlist
will be called on all ranks, but only those cases targeted to
this rank will run. Override this method
(see LatinHypercubeDriver) if your DOE generator needs to
create all cases on one rank and scatter them to other ranks.
"""
for i, case in enumerate(self._build_runlist()):
if (i % self._num_par_doe) == self._par_doe_id:
yield case
def _distrib_lb_build_runlist(self):
"""
Runs a load balanced version of the runlist, with the master
rank (0) sending a new case to each worker rank as soon as it
has finished its last case.
"""
comm = self._full_comm
if self._full_comm.rank == 0: # master rank
runiter = self._build_runlist()
received = 0
sent = 0
# cases left for each par doe
cases = {n:{'count': 0, 'p':{}, 'u':{}, 'r':{}, 'meta':{}}
for n in self._id_map}
# create a mapping of ranks to doe_ids
doe_ids = {}
for doe_id, tup in self._id_map.items():
size, offset = tup
for i in range(size):
doe_ids[i+offset] = doe_id
# seed the workers
for i in range(1, self._num_par_doe):
try:
# case is a generator, so must make a list to send
case = list(next(runiter))
except StopIteration:
break
size, offset = self._id_map[i]
for j in range(size):
comm.send(case, j+offset, tag=1)
cases[i]['count'] += 1
sent += 1
# send the rest of the cases
if sent > 0:
more_cases = True
while True:
worker, p, u, r, meta = comm.recv()
received += 1
caseinfo = cases[doe_ids[worker]]
caseinfo['count'] -= 1
caseinfo['p'].update(p)
caseinfo['u'].update(u)
caseinfo['r'].update(r)
caseinfo['meta'].update(meta)
if caseinfo['count'] == 0:
# we've received case from all procs with that doe_id
# so the case is complete. Send it to recorders
yield caseinfo
if more_cases:
try:
case = list(next(runiter))
except StopIteration:
more_cases = False
else:
size, offset = self._id_map[doe_ids[worker]]
for j in range(size):
comm.send(case, j+offset, tag=1)
cases[doe_ids[worker]]['count'] += 1
sent += 1
if received == sent:
break
# tell all workers to stop
for rank in range(1, self._full_comm.size):
comm.isend(None, rank, tag=1)
else: # worker
while True:
# wait on a case from the master
case = comm.recv(source=0, tag=1)
if case is None: # we're done
break
# yield the case so it can be executed
yield case
# get local vars from RecordingManager
params, unknowns, resids = self.recorders._get_local_case_data(self.root)
# tell the master we're done with that case and send local vars
comm.send((comm.rank, params, unknowns, resids, self._last_meta), 0)
