"""
Baseclass for design-of-experiments Drivers that have pre-determined
parameter sets.
"""
from __future__ import print_function
import sys
import os
import traceback
import logging
from six.moves import zip
from six import next, PY3
import multiprocessing
import numpy
from openmdao.core.problem import _get_root_var
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, any_proc_is_true
from openmdao.core.system import AnalysisError
trace = os.environ.get('OPENMDAO_TRACE')
[docs]def worker(problem, response_vars, case_queue, response_queue, worker_id):
"""This is used to run parallel DOEs using multprocessing. It takes a case
off of the case_queue, runs it, then puts responses on the response_queue.
"""
# set env var so comps/recorders know they're running in a worker proc
os.environ['OPENMDAO_WORKER_ID'] = str(worker_id)
try:
# on windows all of our args are pickled, which causes us to lose the
# connections between our numpy views and their parent arrays, so force
# the problem to setup() again.
if sys.platform == 'win32':
problem.setup(check=False)
driver = problem.driver
root = driver.root
terminate = 0
for case_id, case in iter(case_queue.get, 'STOP'):
#logging.info("worker %d, case id %d, case %s" % (worker_id, case_id, case))
if terminate:
continue
metadata = driver._prep_case(case, case_id)
try:
terminate, exc = driver._try_case(root, metadata)
if terminate:
complete_case = (metadata, [])
else:
complete_case = (metadata,
[_get_root_var(root, n) for n in response_vars])
except:
# we generally shouldn't get here, but just in case,
# handle it so that the main process doesn't hang at the
# end when it tries to join all of the concurrent processes.
if metadata.get('msg'):
metadata['msg'] += "\n\n%s" % traceback.format_exc()
else:
metadata['msg'] = traceback.format_exc()
metadata['success'] = 0
metadata['terminate'] = 1
complete_case = (metadata, [])
metadata['id'] = case_id
response_queue.put(complete_case)
except:
logging.error(traceback.format_exc())
raise
[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 and running under MPI, use rank 0 as master and load balance
cases among all of the other ranks. Default is False. If
multiprocessing is being used instead of MPI, then cases are always
load balanced.
"""
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 = int(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:
self._num_par_doe = 1
self._load_balance = False
self._full_comm = comm
# figure out which parallel DOE we are associated with
if MPI and 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()
if MPI:
minprocs *= self._num_par_doe
if MPI and 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
with problem.root._dircontext:
if self._num_par_doe > 1:
if MPI:
if self._load_balance:
self._run_lb(problem.root)
else:
self._run_par_doe(problem.root)
else: # use multiprocessing
self._run_lb_multiproc(problem)
else:
self._run_serial()
def _save_case(self, case, meta=None):
if self._num_par_doe > 1:
if self._load_balance:
self.recorders.record_completed_case(self.root, case)
else:
self.recorders.record_iteration(self.root, meta,
dummy=(case is None))
else:
self.recorders.record_iteration(self.root, meta)
def _prep_case(self, case, iter_count):
"""Create metadata for the case and set design variables.
"""
metadata = create_local_meta(None, 'Driver')
update_local_meta(metadata, (iter_count,))
for dv_name, dv_val in case:
self.set_desvar(dv_name, dv_val)
return metadata
def _try_case(self, root, metadata):
"""Run a case and save exception info and mark the metadata
if the case fails.
"""
terminate = False
exc = None
metadata['terminate'] = 0
try:
root.solve_nonlinear(metadata=metadata)
except AnalysisError:
metadata['msg'] = traceback.format_exc()
metadata['success'] = 0
except Exception:
metadata['success'] = 0
# this will tell master to stop sending cases in lb case
metadata['terminate'] = 1
metadata['msg'] = traceback.format_exc()
print(metadata['msg'])
if not self._load_balance:
exc = sys.exc_info()
terminate = True
return terminate, exc
def _run_serial(self):
"""This runs a DOE in serial on a single process."""
root = self.root
for case in self._build_runlist():
metadata = self._prep_case(case, self.iter_count)
terminate, exc = self._try_case(root, metadata)
if exc is not None:
if PY3:
raise exc[0].with_traceback(exc[1], exc[2])
else:
# exec needed here since otherwise python3 will
# barf with a syntax error :(
exec('raise exc[0], exc[1], exc[2]')
self._save_case(case, metadata)
self.iter_count += 1
def _run_par_doe(self, root):
"""This runs the DOE in parallel where cases are evenly distributed
among all processes.
"""
for case in self._get_case_w_nones(self._distrib_build_runlist()):
if case is None: # dummy cases have case == None
# must take part in collective Allreduce call
any_proc_is_true(self._full_comm, False)
else: # case is not a dummy case
metadata = self._prep_case(case, self.iter_count)
terminate, exc = self._try_case(root, metadata)
if any_proc_is_true(self._full_comm, terminate):
if exc:
if PY3:
raise exc[0].with_traceback(exc[1], exc[2])
else:
# exec needed here since otherwise python3 will
# barf with a syntax error :(
exec('raise exc[0], exc[1], exc[2]')
else:
raise RuntimeError("an exception was raised by another MPI process.")
self._save_case(case, metadata)
self.iter_count += 1
def _run_lb(self, root):
"""This runs the DOE in parallel with load balancing via MPI. A new case
is distributed to a worker process as soon as it finishes its
previous case. The rank 0 process is the 'master' process and does
not run cases itself. The master does nothing but distribute the
cases to the workers and collect the results.
"""
for case in self._distrib_lb_build_runlist():
if self._full_comm.rank == 0:
# we're the master rank and case is a completed case
self._save_case(case)
else: # we're a worker
metadata = self._prep_case(case, self.iter_count)
self._try_case(root, metadata)
# keep meta for worker to send to master
self._last_meta = metadata
self.iter_count += 1
def _build_case(self, meta, uvars, pvars, numuvars, values):
"""
Given values returned from a multiproc run, construct
a case object that can be passed to recorders.
"""
if meta['terminate']:
print("Worker has requested termination. No more new "
"cases will be distributed. Worker traceback was:\n%s" %
meta['msg'])
return None
return {
'u':{n:v for n,v in zip(uvars, values)},
'p':{n:v for n,v in zip(pvars, values[numuvars:])},
'r':{},
'meta': meta
}
def _run_lb_multiproc(self, problem):
"""This runs the DOE in parallel with load balancing via
multiprocessing. A new case is distributed to a worker process as
soon as it finishes its previous case.
"""
root = problem.root
uvars = list(self.recorders._vars_to_record['unames'])
pvars = list(self.recorders._vars_to_record['pnames'])
response_vars = uvars + pvars
numuvars = len(uvars)
runiter = self._build_runlist()
# Create queues
if sys.platform == 'win32':
manager = multiprocessing.Manager()
task_queue = manager.Queue()
done_queue = manager.Queue()
else:
task_queue = multiprocessing.Queue()
done_queue = multiprocessing.Queue()
procs = []
terminating = False
# Start worker processes
for i in range(self._num_par_doe):
procs.append(multiprocessing.Process(target=worker,
args=(problem, response_vars,
task_queue, done_queue, i)))
for proc in procs:
proc.start()
iter_count = 0
num_active = 0
empty = {}
try:
for proc in procs:
# case is a generator, so must make a list to send
case = list(next(runiter))
task_queue.put((iter_count, case))
iter_count += 1
num_active += 1
except StopIteration:
pass
else:
try:
while num_active > 0:
meta, values = done_queue.get()
#logging.info("RECEIVED: %d, %s" % (meta['id'], values[2]))
complete_case = self._build_case(meta, uvars, pvars,
numuvars, values)
num_active -= 1
if complete_case is None:
# there was a fatal error, don't run more cases
break
self.recorders.record_completed_case(root, complete_case)
case = list(next(runiter))
task_queue.put((iter_count, case))
iter_count += 1
num_active += 1
except StopIteration:
pass
# tell all workers we're done
for proc in procs:
task_queue.put('STOP')
for i in range(num_active):
meta, values = done_queue.get()
#logging.info("RECEIVED: %d, %s" % (meta['id'], values[0]))
complete_case = self._build_case(meta, uvars, pvars,
numuvars, values)
if complete_case is None:
# had a fatal error, don't record
continue
self.recorders.record_completed_case(root, complete_case)
for proc in procs:
proc.join()
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, 'terminate': 0, 'p':{}, 'u':{}, 'r':{},
'meta':{'success': 1, 'msg': ''}}
for n in self._id_map}
# create a mapping of ranks to doe_ids, to handle those cases
# where a single DOE is executed across multiple processes, i.e.,
# for each process, we need to know which case it's working on.
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]
# send the case to all of the subprocs that will work on it
for j in range(size):
if trace:
debug('Sending Seed case %d, %d' % (i, j))
comm.send(case, j+offset, tag=1)
if trace:
debug('Seed Case Sent %d, %d' % (i, j))
cases[i]['count'] += 1
sent += 1
# send the rest of the cases
if sent > 0:
more_cases = True
while True:
if trace: debug("Waiting on case")
worker, p, u, r, meta = comm.recv(tag=2)
if trace: debug("Case Recieved from Worker %d" % worker )
received += 1
caseinfo = cases[doe_ids[worker]]
caseinfo['count'] -= 1
caseinfo['p'].update(p)
caseinfo['u'].update(u)
caseinfo['r'].update(r)
# save certain parts of existing metadata so we don't hide failures
oldmeta = caseinfo['meta']
success = oldmeta['success']
if not success:
msg = oldmeta['msg']
oldmeta.update(meta)
oldmeta['success'] = success
oldmeta['msg'] = msg
else:
oldmeta.update(meta)
caseinfo['terminate'] += meta.get('terminate', 0)
if caseinfo['count'] == 0:
# we've received case from all procs with that doe_id
# so the case is complete.
# worker has experienced some critical error, so we'll
# stop sending new cases and start to wrap things up
if caseinfo['terminate'] > 0:
more_cases = False
print("Worker %d has requested termination. No more new "
"cases will be distributed. Worker traceback was:\n%s" %
(worker, meta['msg']))
else:
# Send case to recorders
yield caseinfo
if more_cases:
try:
case = list(next(runiter))
except StopIteration:
more_cases = False
else:
# send a new case to every proc that works on
# cases with the current worker
doe = doe_ids[worker]
size, offset = self._id_map[doe]
cases[doe]['terminate'] = 0
cases[doe]['meta'] = {'success': 1, 'msg': ''}
for j in range(size):
if trace:
debug("Sending New Case to Worker %d" % worker )
comm.send(case, j+offset, tag=1)
if trace:
debug("Case Sent to Worker %d" % worker )
cases[doe]['count'] += 1
sent += 1
# don't stop until we hear back from every worker process
# we sent a case to
if received == sent:
break
# tell all workers to stop
for rank in range(1, self._full_comm.size):
if trace:
debug("Make Worker Stop on Rank %d" % rank )
comm.send(None, rank, tag=1)
if trace:
debug("Worker has Stopped on Rank %d" % rank )
else: # worker
while True:
# wait on a case from the master
if trace: debug("Receiving Case from Master")
case = comm.recv(source=0, tag=1)
if trace: debug("Case Received from Master")
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
if trace: debug("Send Master Local Vars")
comm.send((comm.rank, params, unknowns, resids, self._last_meta), 0, tag=2)
if trace: debug("Local Vars Sent to Master")
