def scatter_dict(whole):
"""
Broadcast and recieve a dictionary where the values are 1d arrays
and the arrays are chunked for the workers.
Only rank 0 needs the whole dictionary.
:param whole: The dictionary of 1d arrays to subdict.
:returns: (chunk of dictionary of 1d arrays, indexes of whole array)
"""
if not STATE.is_parallel:
array_len = len(whole[whole.keys()[0]])
return whole, numpy.array(range(0, array_len))
else:
import pypar # pylint: disable=W0404
if STATE.rank == 0:
array_len = len(whole[whole.keys()[0]])
for pro in range(0, STATE.size):
temp_indexes = numpy.array(range(pro, array_len, STATE.size))
temp_subdict = {}
for key in whole.keys():
temp_subdict[key] = whole[key][temp_indexes]
if pro is 0:
indexes = temp_indexes
subdict = temp_subdict
else:
pypar.send(temp_indexes, pro)
pypar.send(temp_subdict, pro)
else:
indexes = pypar.receive(0)
subdict = pypar.receive(0)
return subdict, indexes
def get_enquiry_depths(self):
# Should be called from all processors associated with operator
enq0 = None
enq1 = None
get0 = "self.inlets[0].get_enquiry_depth()"
get1 = "self.inlets[1].get_enquiry_depth()"
if self.myid == self.master_proc:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
else:
enq0 = pypar.receive(self.enquiry_proc[0])
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
else:
enq1 = pypar.receive(self.enquiry_proc[1])
else:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
pypar.send(enq0, self.master_proc)
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
pypar.send(enq1, self.master_proc)
return [enq0, enq1]
def get_enquiry_depths(self):
# Should be called from all processors associated with operator
enq0 = None
enq1 = None
get0 = 'self.inlets[0].get_enquiry_depth()'
get1 = 'self.inlets[1].get_enquiry_depth()'
if self.myid == self.master_proc:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
else:
enq0 = pypar.receive(self.enquiry_proc[0])
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
else:
enq1 = pypar.receive(self.enquiry_proc[1])
else:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
pypar.send(enq0, self.master_proc)
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
pypar.send(enq1, self.master_proc)
return [enq0, enq1]
def get_enquiry_water_depths(self):
enq0 = None
enq1 = None
get0 = 'self.inlets[0].get_enquiry_water_depth()'
get1 = 'self.inlets[1].get_enquiry_water_depth()'
if self.myid == self.master_proc:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
else:
enq0 = pypar.receive(self.enquiry_proc[0])
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
else:
enq1 = pypar.receive(self.enquiry_proc[1])
else:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
pypar.send(enq0, self.master_proc)
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
pypar.send(enq1, self.master_proc)
return [enq0, enq1]
def scatter_dict(whole):
"""
Broadcast and recieve a dictionary where the values are 1d arrays
and the arrays are chunked for the workers.
Only rank 0 needs the whole dictionary.
:param whole: The dictionary of 1d arrays to subdict.
:returns: (chunk of dictionary of 1d arrays, indexes of whole array)
"""
if not STATE.is_parallel:
array_len = len(whole[list(whole.keys())[0]])
return whole, numpy.array(list(range(0, array_len)))
else:
import pypar # pylint: disable=W0404
if STATE.rank == 0:
array_len = len(whole[list(whole.keys())[0]])
for pro in range(0, STATE.size):
temp_indexes = numpy.array(list(range(pro, array_len, STATE.size)))
temp_subdict = {}
for key in list(whole.keys()):
temp_subdict[key] = whole[key][temp_indexes]
if pro == 0:
indexes = temp_indexes
subdict = temp_subdict
else:
pypar.send(temp_indexes, pro)
pypar.send(temp_subdict, pro)
else:
indexes = pypar.receive(0)
subdict = pypar.receive(0)
return subdict, indexes
def balanceArrays(self, arrayFragment):
'''Redistributes the elements in a set of arrays equally across the nodes'''
if Environment.isParallel:
import pypar
if self.isRoot():
completeArray = arrayFragment
for i in range(1, pypar.size()):
fragment = pypar.receive(i)
completeArray.extend(fragment)
#Divide it up
divisions = self._divideArray(completeArray)
#Send the fragments
for i in range(1, pypar.size()):
start, end = divisions[i]
pypar.send(completeArray[start:end], i)
self.output('[ENV] Rebalanced array divisions %s' % divisions)
#Assign root fragment
start, end = divisions[0]
arrayFragment = completeArray[start:end]
else:
#Send the fragment
pypar.send(arrayFragment, 0)
#Retrieve the array
arrayFragment = pypar.receive(0)
else:
completeArray = arrayFragment
return arrayFragment
def get_enquiry_water_depths(self):
enq0 = None
enq1 = None
get0 = 'self.inlets[0].get_enquiry_water_depth()'
get1 = 'self.inlets[1].get_enquiry_water_depth()'
if self.myid == self.master_proc:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
else:
enq0 = pypar.receive(self.enquiry_proc[0])
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
else:
enq1 = pypar.receive(self.enquiry_proc[1])
else:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
pypar.send(enq0, self.master_proc)
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
pypar.send(enq1, self.master_proc)
return [enq0, enq1]
def get_enquiry_velocitys(self):
enq0 = None
enq1 = None
get0 = "self.inlets[0].get_enquiry_velocity()"
get1 = "self.inlets[1].get_enquiry_velocity()"
if self.myid == self.master_proc:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
else:
enq0 = pypar.receive(self.enquiry_proc[0])
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
else:
enq1 = pypar.receive(self.enquiry_proc[1])
else:
if self.myid == self.enquiry_proc[0]:
enq0 = eval(get0)
pypar.send(enq0, self.master_proc)
if self.myid == self.enquiry_proc[1]:
enq1 = eval(get1)
pypar.send(enq1, self.master_proc)
return [enq0, enq1]
def multi_proc_mutate_and_integrate(self, prmt, mutation):
"""subroutine to send jobs to multiple processors using pypar
It send one network - and not the complete population- to each processor
and explicitly we take care of synchronization.
Args:
prmt (dict): the inits parameters for integration
mutation (list): tuple (id,mut) indicating the mutation flag of integration for each network
Returns:
int: the total number of mutations
"""
numproc = self.numproc #computes number of available proc for integration : the proc 0 is used as master proc
l = len(mutation)
n_mut = 0
for index_job in range(l):
nproc = 1 + index_job % (
numproc - 1
) #computes the proc number where to send the job, we start at 1, proc 0 is master proc
args = {
'net': self.genus[index_job],
'prmt': prmt,
'nnetwork': index_job,
'tgeneration': self.tgeneration,
'mutation': mutation[index_job]
}
pypar.send((
'net.mutate_and_integrate(prmt,nnetwork,tgeneration,mutation)',
args), nproc) #send integration job to the selected proc
if ((index_job + 1) % (numproc - 1) == 0):
pypar.barrier(
) # every numproc jobs sent, we wait for all other processors to finish their job to synchronize
results = [
pypar.receive(worker) for worker in range(1, numproc)
] #receives results from all processors
for i in results:
n_mut += i[0] #updates number of mutations
self.genus[i[1]] = i[2] #updates mutated network
self.update_fitness(i[1], i[3]) #updates fitness values
#at that point, we may have jobs still running on some subset of the procs, but we only take the results from the last working processors
if (l % (numproc - 1) > 0):
for i in range(l % (numproc - 1), numproc - 1):
pypar.send(
('0', {}), i + 1
) #send dummy jobs to the still processors for synchornization purpose
pypar.barrier() #synchronize the proc
results = [pypar.receive(worker) for worker in range(1, numproc)]
#only take the results we are interested in
for i in range(l % (numproc - 1)):
n_mut += results[i][0]
self.genus[results[i][1]] = results[i][2]
self.update_fitness(results[i][1],
results[i][3]) #updates fitness values
# Shut down workers
#for worker in range(1,numproc):
# pypar.send(SystemExit(),worker)
return n_mut
def master(self):
self.numCompleted = 0
self.mapList = list()
logging.info('[MASTER]: started processor %d of %d on node %s: number of works: %d'%(self.MPI_myid, self.MPI_numproc, self.MPI_node, self.numWorks))
# start slaves distributing the first work slot
rounder = 0
if self.MPI_numproc <= self.numWorks:
rounder = 1
for i in range(min(self.MPI_numproc, self.numWorks)-rounder):
work = self.works[i]
pypar.send(work, destination=i+1, tag=self.WORKTAG)
logging.debug('[MASTER]: sent work "%s" to node %d' %(work, i+1))
# dispatch the remaining work slots on dynamic load-balancing policy
# the quicker to do the job, the more jobs it takes
for work in self.works[self.MPI_numproc-1:]:
result, status = pypar.receive(source=pypar.any_source, tag=self.WORKTAG,
return_status=True)
logging.debug('[MASTER]: received result "%s" from node %d'%(result, status.source))
self.mapList.append(result)
self.numCompleted += 1
logging.debug('[MASTER]: done : %d' %self.numCompleted)
pypar.send(work, destination=status.source, tag=self.WORKTAG)
logging.debug('[MASTER]: sent work "%s" to node %d' %(work, status.source))
# all works have been dispatched out
logging.debug('[MASTER]: toDo : %d' %self.numWorks)
logging.debug('[MASTER]: done : %d' %self.numCompleted)
# I've still to take into the remaining completions
while (self.numCompleted < self.numWorks):
result, status = pypar.receive(source=pypar.any_source, tag=self.WORKTAG,
return_status=True)
logging.debug('[MASTER]: received (final) result "%s" from node %d'%(result, status.source))
self.mapList.append(result)
self.numCompleted += 1
logging.debug('[MASTER]: %d completed' %self.numCompleted)
logging.debug('[MASTER]: about to terminate slaves')
# Tell slaves to stop working
for i in range(1, self.MPI_numproc):
pypar.send('#', destination=i, tag=self.DIETAG)
logging.debug('[MASTER]: sent termination signal to node %d' %(i, ))
# call the reduce function
logging.info('[MASTER]: about to run reduce')
res = self.reduceFunction(self.mapList)
return res
def master(self):
numcompleted = 0
#--- start slaves distributing the first work slot
for i in range(0, min(self.numprocs - 1, self.numworks)):
work = i
slave = i + 1
pypar.send(work, destination=slave, tag=PYPAR_WORKTAG)
if self.debug:
print '[MASTER ]: sent first work "%s" to node %d' % (work,
slave)
# dispatch the remaining work slots on dynamic load-balancing policy
# the quicker to do the job, the more jobs it takes
for work in range(self.numprocs - 1, self.numworks):
result, status = pypar.receive(source=pypar.any_source,
tag=PYPAR_WORKTAG,
return_status=True)
if self.debug:
print '[MASTER ]: received result from node %d' % (
status.source, )
numcompleted += 1
pypar.send(work, destination=status.source, tag=PYPAR_WORKTAG)
if self.debug:
print '[MASTER ]: sent work "%s" to node %d' % (work,
status.source)
self.work.handleWorkResult(result, status)
# all works have been dispatched out
if self.debug: print '[MASTER ]: ToDo : %d' % self.numworks
if self.debug: print '[MASTER ]: Done : %d' % numcompleted
# I've still to take into the remaining completions
while (numcompleted < self.numworks):
result, status = pypar.receive(source=pypar.any_source,
tag=PYPAR_WORKTAG,
return_status=True)
if self.debug:
print '[MASTER ]: received (final) result from node %d' % (
status.source, )
numcompleted += 1
if self.debug: print '[MASTER ]: %d completed' % numcompleted
self.work.handleWorkResult(result, status)
if self.debug: print '[MASTER ]: about to terminate slaves'
# Tell slaves to stop working
for i in range(1, self.numprocs):
pypar.send('#', destination=i, tag=PYPAR_DIETAG)
if self.debug: print '[MASTER ]: sent DIETAG to node %d' % (i, )
def mpi_sync_db(session):
"""Causes the halo_db module to use the rank 0 processor's 'Creator' object"""
global _mpi_initialized
if _mpi_initialized:
import pypar
import halo_db as db
if pypar.rank() == 0:
x = session.merge(db._current_creator)
session.commit()
time.sleep(0.5)
print("Manager --> transmit run ID=", x.id)
for i in range(1, pypar.size()):
pypar.send(x.id, tag=3, destination=i)
db._current_creator = x
else:
ID = pypar.receive(source=0, tag=3)
print("Rank", pypar.rank(), " --> set run ID=", ID)
db._current_creator = session.query(
db.Creator).filter_by(id=ID).first()
print(db._current_creator)
else:
print("NOT syncing DB references: MPI unavailable")
def test_mpi_recvsend(target_array, message_range, target_rank, tag_mark):
send_range = message_range[0]
recv_range = message_range[1]
target_array[recv_range] = mpi.receive(target_rank,tag=tag_mark)
# print 'I`m', myrank, 'Recv : ', target_rank, 'range : ', recv_range
mpi.send(target_array[send_range].copy(), target_rank, tag=tag_mark)
def slave(self):
if self.debug:
print '[SLAVE %d]: I am processor %d of %d on node %s' % (
self.myid, self.myid, self.numprocs, self.node)
if self.debug: print '[SLAVE %d]: Entering work loop' % (self.myid, )
while True:
result, status = pypar.receive(source=0,
tag=pypar.any_tag,
return_status=True)
print '[SLAVE %d]: received work with tag %d from node %d'\
%(self.myid, status.tag, status.source)
if (status.tag == PYPAR_DIETAG):
print '[SLAVE %d]: received termination from node %d' % (
self.myid, 0)
return
else:
worknum = result
if self.debug:
print '[SLAVE %d]: work number is %s' % (self.myid,
worknum)
myresult = self.work.calcWorkResult(worknum)
pypar.send(myresult, destination=0)
if self.debug:
print '[SLAVE %d]: sent result to node %d' % (self.myid, 0)
def splitArray(self, array): '''Splits array between all the processes in the environment. Each process will be returned a different section of the array to work on''' if Environment.isParallel: import pypar #Split the array into sections and return the section for this processor divisions = [] if self.isRoot(): #Root does the splitting - we send each processor the start and end index #NOTE: pypar broadcast won't work even when setting vanilla #It always returns message trucated error. divisions = self._divideArray(array) for i in range(1,pypar.size()): pypar.send(divisions[i], i) start = divisions[0][0] end = divisions[0][1] else: indexes = pypar.receive(0) start = indexes[0] end = indexes[1] return array[start:end] else: return array
def receive_any(source=None):
if source is None:
source = pypar.any_source
data, status = pypar.receive(source=source,
return_status=True,
tag=pypar.any_tag)
return data, status.source, status.tag
def broadcast( obj ): if root(): for i in xrange(p.size()-1): p.send(obj,i+1) return obj else: return p.receive(0)
def scatter(vec):
if root():
for i in xrange(p.size() - 1):
p.send(vec[i + 1], i + 1)
return vec[0]
else:
return p.receive(0)
def broadcast(obj):
if root():
for i in xrange(p.size() - 1):
p.send(obj, i + 1)
return obj
else:
return p.receive(0)
def distributed_generator(iterable):
"""
Distribute the values from a generator across workers.
"""
RUN, DIE = range(2)
P = pp.size()
if P == 1:
for el in iterable:
yield el
else:
if pp.rank() == 0:
it = iter(iterable)
while True:
try:
first = next(it)
for p in range(1, P):
pp.send(next(it), p, tag=RUN)
yield first
except StopIteration:
for p in range(1, P):
pp.send(666, p, tag=DIE)
break
else:
while True:
el, status = pp.receive(0, tag=pp.any_tag, return_status=True)
if status.tag == DIE:
break
yield el
def slave(self):
logging.debug(
'[SLAVE %d]: started processor %d of %d on node %s' %
(self.MPI_myid, self.MPI_myid, self.MPI_numproc, self.MPI_node))
while True:
inputMsg, status = pypar.receive(source=0,
tag=pypar.any_tag,
return_status=True)
logging.debug(
'[SLAVE %d]: received work "%s" with tag %d from node %d' %
(self.MPI_myid, inputMsg, status.tag, status.source))
if (status.tag == self.DIETAG):
logging.debug('[SLAVE %d]: received termination from node %d' %
(self.MPI_myid, 0))
return
else:
logging.debug('[SLAVE %d]: received work "%s" to map' %
(self.MPI_myid, inputMsg))
resultMsg = self.mapFunction(inputMsg)
pypar.send(resultMsg, destination=0, tag=self.WORKTAG)
logging.debug('[SLAVE %d]: sent result "%s" to node %d' %
(self.MPI_myid, resultMsg, 0))
def test_longint_array(self):
myid, ncpu = pp.rank(), pp.size()
N = 17 # Number of elements
if myid == 0:
A = np.array(range(N)).astype('l')
B = np.zeros(N).astype('l')
pp.send(A, 1, use_buffer=True)
B = pp.receive(ncpu - 1, buffer=B)
self.assertTrue(np.allclose(A, B))
else:
X = np.zeros(N).astype('l')
X = pp.receive(myid - 1, buffer=X)
pp.send(X, (myid + 1) % ncpu, use_buffer=True)
def scatter( vec ): if root(): for i in xrange(p.size()-1): p.send(vec[i+1],i+1) return vec[0] else: return p.receive(0)
def _mpi_assign_thread(job_iterator):
# Sit idle until request for a job comes in, then assign first
# available job and move on. Jobs are labelled through the
# provided iterator
import pypar
import pypar.mpiext
j = -1
print("Manager --> Entered iterator code")
alive = [True for i in range(pypar.size())]
while any(alive[1:]):
dest = pypar.receive(source=pypar.mpiext.MPI_ANY_SOURCE, tag=1)
try:
time.sleep(0.05)
j = job_iterator.next()[0]
print("Manager --> Sending job", j, "to rank", dest)
except StopIteration:
alive[dest] = False
print("Manager --> Sending out of job message to ", dest)
j = None
pypar.send(j, destination=dest, tag=2)
print("Manager --> All jobs done and all processors>0 notified; exiting thread")
def get_global_average_elevation(self):
# GLOBAL: Master processor gathers elevations from all child processors, and returns average
# WARNING: requires synchronization, must be called by all procs associated
# with this inlet
import pypar
local_elevation = num.sum(self.get_elevations() * self.get_areas())
global_area = self.get_global_area()
global_elevation = local_elevation
if self.myid == self.master_proc:
for i in self.procs:
if i == self.master_proc: continue
val = pypar.receive(i)
global_elevation = global_elevation + val
else:
pypar.send(local_elevation, self.master_proc)
if global_area > 0.0:
return global_elevation / global_area
else:
return 0.0
def get_global_average_elevation(self):
# GLOBAL: Master processor gathers elevations from all child processors, and returns average
# WARNING: requires synchronization, must be called by all procs associated
# with this inlet
import pypar
local_elevation = num.sum(self.get_elevations()*self.get_areas())
global_area = self.get_global_area()
global_elevation = local_elevation
if self.myid == self.master_proc:
for i in self.procs:
if i == self.master_proc: continue
val = pypar.receive(i)
global_elevation = global_elevation + val
else:
pypar.send(local_elevation, self.master_proc)
if global_area > 0.0:
return global_elevation/global_area
else:
return 0.0
def test_longint_array(self):
myid, ncpu = pp.rank(), pp.size()
N = 17 # Number of elements
if myid == 0:
A = np.array(range(N)).astype('l')
B = np.zeros(N).astype('l')
pp.send(A, 1, use_buffer=True)
B = pp.receive(ncpu - 1, buffer=B)
self.assertTrue(np.allclose(A, B))
else:
X = np.zeros(N).astype('l')
X = pp.receive(myid - 1, buffer=X)
pp.send(X, (myid + 1) % ncpu, use_buffer=True)
def master():
numCompleted = 0
print '[MASTER]: I am processor %d of %d on node %s'\
%(MPI_myid, MPI_numproc, MPI_node)
# start slaves distributing the first work slot
for i in range(1, min(MPI_numproc, numWorks)):
work = workList[i]
pypar.send(work, destination=i, tag=WORKTAG)
print '[MASTER]: sent work "%s" to node %d' %(work, i)
# dispatch the remaining work slots on dynamic load-balancing policy
# the quicker to do the job, the more jobs it takes
for work in workList[MPI_numproc:]:
result, status = pypar.receive(source=pypar.any_source, tag=WORKTAG,
return_status=True)
print '[MASTER]: received result "%s" from node %d'\
%(result, status.source)
numCompleted += 1
pypar.send(work, destination=status.source, tag=WORKTAG)
print '[MASTER]: sent work "%s" to node %d' %(work, status.source)
# all works have been dispatched out
print '[MASTER]: toDo : %d' %numWorks
print '[MASTER]: done : %d' %numCompleted
# I've still to take into the remaining completions
while (numCompleted < numWorks):
result, status = pypar.receive(source=pypar.any_source,
tag=WORKTAG,
return_status=True)
print '[MASTER]: received (final) result "%s" from node %d'\
%(result, status.source)
numCompleted += 1
print '[MASTER]: %d completed' %numCompleted
print '[MASTER]: about to terminate slaves'
# Tell slaves to stop working
for i in range(1, MPI_numproc):
pypar.send('#', destination=i, tag=DIETAG)
print '[MASTER]: sent termination signal to node %d' %(i, )
return
def Calculate(self, varsByCalc, params=None):
"""
Calculate model predictions for everything in varsByCalc.
varsByCalc is a dictionary of the form:
dict[calc name][dep var] = ind var
The return dictionary is of the form:
dictionary[calc name][dep var][ind var] = result
"""
if params is not None:
self.params.update(params)
results = {}
calcs_to_do = varsByCalc.keys()
# Record which calculation each node is doing
calc_assigned = {}
while calcs_to_do:
# The number of calculations to do this round. We want to use
# all the processors if possible.
len_this_block = min(SloppyCell.num_procs, len(calcs_to_do))
for worker in range(1, len_this_block):
calc = calcs_to_do.pop()
calc_assigned[worker] = calc
logger.debug('Assigning calculation %s to worker %i.' %
(calc, worker))
command = 'Network.calculate(net, vars, params)'
args = {
'net': self.get(calc),
'vars': varsByCalc[calc],
'params': self.params
}
pypar.send((command, args), worker)
# The master does his share here
calc = calcs_to_do.pop()
# We use the finally statement because we want to ensure that we
# *always* wait for replies from the workers, even if the master
# encounters an exception in his evaluation.
try:
results[calc] = self.get(calc).calculate(
varsByCalc[calc], self.params)
finally:
# Collect results from the workers
for worker in range(1, len_this_block):
logger.debug('Receiving result from worker %i.' % worker)
results[calc_assigned[worker]] = pypar.receive(worker)
# If the master encounts an exception, we'll break out of the
# function ***here***
# Check the results we received. If any is a SloppyCellException,
# reraise it.
for val in results.values():
if isinstance(val, Utility.SloppyCellException):
raise val
return results
def statistics(self):
# Warning: requires synchronization, must be called by all procs associated
# with this structure
message = ' '
if self.myid == self.master_proc:
message = '===============================================\n'
message += 'Parallel Structure Operator: %s\n' % self.label
message += '===============================================\n'
message += 'Structure Type: %s\n' % self.structure_type
message += 'Description\n'
message += '%s' % self.description
message += '\n'
#add the culvert dimensions, blockage factor here
if self.structure_type == 'boyd_pipe':
message += 'Culvert Diameter: %s\n' % self.diameter
message += 'Culvert Blockage: %s\n' % self.blockage
message += 'No. of barrels: %s\n' % self.barrels
elif self.structure_type == 'boyd_box':
message += 'Culvert Height: %s\n' % self.height
message += 'Culvert Width: %s\n' % self.width
message += 'Culvert Blockage: %s\n' % self.blockage
message += 'No. of barrels: %s\n' % self.barrels
else:
message += 'Culvert Height: %s\n' % self.height
message += 'Culvert Width: %s\n' % self.width
message += 'Batter Slope 1: %s\n' % self.z1
message += 'Batter Slope 2: %s\n' % self.z2
#print "Structure Myids ",self.myid, self.label
for i, inlet in enumerate(self.inlets):
if self.myid == self.master_proc:
message += '-------------------------------------\n'
message += 'Inlet %i\n' % (i)
message += '-------------------------------------\n'
#print "*****",inlet, i,self.myid
if inlet is not None:
stats = inlet.statistics()
if self.myid == self.master_proc:
if self.myid != self.inlet_master_proc[i]:
stats = pypar.receive(self.inlet_master_proc[i])
elif self.myid == self.inlet_master_proc[i]:
pypar.send(stats, self.master_proc)
if self.myid == self.master_proc: message += stats
if self.myid == self.master_proc:
message += '=====================================\n'
return message
def broadcast_vec( vec, i ): myid = p.rank() if myid == i: for j in xrange(p.size()): if j != myid: p.send(vec[i],j) else: vec[i] = p.receive(i)
def broadcast_vec(vec, i):
myid = p.rank()
if myid == i:
for j in xrange(p.size()):
if j != myid:
p.send(vec[i], j)
else:
vec[i] = p.receive(i)
def gather( obj ): if root(): result = [ None for i in xrange(p.size()) ] result[0] = obj for i in xrange(p.size()-1): result[i+1] = p.receive(i+1) return result else: p.send(obj,0)
def receive(self, *args, **kwargs):
"""
Wrapper for pypar.receive
"""
if self.is_parallel is True:
import pypar
return pypar.receive(*args, **kwargs)
else:
return None
def receive(self, *args, **kwargs):
"""
Wrapper for pypar.receive
"""
if self.is_parallel is True:
import pypar
return pypar.receive(*args, **kwargs)
else:
return None
def gather(obj):
if root():
result = [None for i in xrange(p.size())]
result[0] = obj
for i in xrange(p.size() - 1):
result[i + 1] = p.receive(i + 1)
return result
else:
p.send(obj, 0)
def Calculate(self, varsByCalc, params = None):
"""
Calculate model predictions for everything in varsByCalc.
varsByCalc is a dictionary of the form:
dict[calc name][dep var] = ind var
The return dictionary is of the form:
dictionary[calc name][dep var][ind var] = result
"""
if params is not None:
self.params.update(params)
results = {}
calcs_to_do = varsByCalc.keys()
# Record which calculation each node is doing
calc_assigned = {}
while calcs_to_do:
# The number of calculations to do this round. We want to use
# all the processors if possible.
len_this_block = min(SloppyCell.num_procs, len(calcs_to_do))
for worker in range(1, len_this_block):
calc = calcs_to_do.pop()
calc_assigned[worker] = calc
logger.debug('Assigning calculation %s to worker %i.'
% (calc, worker))
command = 'Network.calculate(net, vars, params)'
args = {'net': self.get(calc), 'vars': varsByCalc[calc],
'params': self.params}
pypar.send((command, args), worker)
# The master does his share here
calc = calcs_to_do.pop()
# We use the finally statement because we want to ensure that we
# *always* wait for replies from the workers, even if the master
# encounters an exception in his evaluation.
try:
results[calc] = self.get(calc).calculate(varsByCalc[calc],
self.params)
finally:
# Collect results from the workers
for worker in range(1, len_this_block):
logger.debug('Receiving result from worker %i.' % worker)
results[calc_assigned[worker]] = pypar.receive(worker)
# If the master encounts an exception, we'll break out of the
# function ***here***
# Check the results we received. If any is a SloppyCellException,
# reraise it.
for val in results.values():
if isinstance(val, Utility.SloppyCellException):
raise val
return results
def __call__(self):
import pypar
volume = 0
# Need to run global command on all processors
current_volume = self.inlet.get_global_total_water_volume()
total_area = self.inlet.get_global_area()
# Only the master proc calculates the update
if self.myid == self.master_proc:
timestep = self.domain.get_timestep()
t = self.domain.get_time()
Q1 = self.update_Q(t)
Q2 = self.update_Q(t + timestep)
volume = 0.5*(Q1+Q2)*timestep
assert current_volume >= 0.0 , 'Volume of watrer in inlet negative!'
for i in self.procs:
if i == self.master_proc: continue
pypar.send((volume, current_volume, total_area, timestep), i)
else:
volume, current_volume, total_area, timestep = pypar.receive(self.master_proc)
#print self.myid, volume, current_volume, total_area, timestep
self.applied_Q = volume/timestep
# Distribute positive volume so as to obtain flat surface otherwise
# just pull water off to have a uniform depth.
if volume >= 0.0 :
self.inlet.set_stages_evenly(volume)
self.domain.fractional_step_volume_integral+=volume
if self.velocity is not None:
# This is done locally without communication
depths = self.inlet.get_depths()
self.inlet.set_xmoms(self.inlet.get_xmoms()+depths*self.velocity[0])
self.inlet.set_ymoms(self.inlet.get_ymoms()+depths*self.velocity[1])
elif current_volume + volume >= 0.0 :
depth = (current_volume + volume)/total_area
self.inlet.set_depths(depth)
self.domain.fractional_step_volume_integral+=volume
else: #extracting too much water!
self.inlet.set_depths(0.0)
self.applied_Q = current_volume/timestep
self.domain.fractional_step_volume_integral-=current_volume
def __call__(self):
import pypar
volume = 0
# Need to run global command on all processors
current_volume = self.inlet.get_global_total_water_volume()
total_area = self.inlet.get_global_area()
# Only the master proc calculates the update
if self.myid == self.master_proc:
timestep = self.domain.get_timestep()
t = self.domain.get_time()
Q1 = self.update_Q(t)
Q2 = self.update_Q(t + timestep)
volume = 0.5 * (Q1 + Q2) * timestep
assert current_volume >= 0.0, 'Volume of watrer in inlet negative!'
for i in self.procs:
if i == self.master_proc: continue
pypar.send((volume, current_volume, total_area, timestep), i)
else:
volume, current_volume, total_area, timestep = pypar.receive(
self.master_proc)
#print self.myid, volume, current_volume, total_area, timestep
self.applied_Q = volume / timestep
# Distribute positive volume so as to obtain flat surface otherwise
# just pull water off to have a uniform depth.
if volume >= 0.0:
self.inlet.set_stages_evenly(volume)
self.domain.fractional_step_volume_integral += volume
if self.velocity is not None:
# This is done locally without communication
depths = self.inlet.get_depths()
self.inlet.set_xmoms(self.inlet.get_xmoms() +
depths * self.velocity[0])
self.inlet.set_ymoms(self.inlet.get_ymoms() +
depths * self.velocity[1])
elif current_volume + volume >= 0.0:
depth = (current_volume + volume) / total_area
self.inlet.set_depths(depth)
self.domain.fractional_step_volume_integral += volume
else: #extracting too much water!
self.inlet.set_depths(0.0)
self.applied_Q = current_volume / timestep
self.domain.fractional_step_volume_integral -= current_volume
def print_l1_stats(full_edge):
numprocs = pypar.size()
myid = pypar.rank()
tri_norm = zeros(3, Float)
recv_norm = zeros(3, Float)
tri_norm[0] = l1_norm(full_edge[:, 0])
tri_norm[1] = l1_norm(full_edge[:, 1])
tri_norm[2] = l1_norm(full_edge[:, 2])
if myid == 0:
for p in range(numprocs-1):
pypar.receive(p+1, recv_norm)
tri_norm[0] = tri_norm[0]+recv_norm[0]
tri_norm[1] = tri_norm[1]+recv_norm[1]
tri_norm[2] = tri_norm[2]+recv_norm[2]
print 'l1_norm along each axis : [', tri_norm[0],', ', tri_norm[1], ', ', tri_norm[2], ']'
else:
pypar.send(tri_norm, 0)
def process(self):
# loop through all the rays that go through the window
for row in xrange(self.Rows):
for col in xrange(self.WindowCols):
# loop through all the light sources
for Light in self.LightSources:
Energy = 0.0
# move the ray to the corresponding pixel on the window. dx and dy changes
MoveUp = self.WindowHeight/2.0 - (row+self.RowStart - 0.5)*\
self.WindowHeight/self.WindowRows # starts from top most pixel
MoveRight = -self.WindowWidth/2.0 + (col + 0.5)*\
self.WindowWidth/self.WindowCols # starts from left most pixel
# define the ray and rotate (in look direction)
# then translate (window distance)
Vec = np.array([self.WindowDistance, -MoveRight, MoveUp])
Vec = self.__RotVec(Vec, self.__Lon, self.__Lat, self.Yaw)
Photon = Ray.Ray(self.LookPos,Vec,1.0)
Cont = True
# get energy after first contact
Res = self.__getEn(Photon, Light)
Energy += Res[0]
UnScatRay = Res[1]
Cont = Res[2]
#pypar.barrier();
for k in range(1):
# get energy after N contacts
if Cont:
Res = self.__getEn(UnScatRay, Light)
Energy += Res[0]
UnScatRay = Res[1]
Cont = Res[2]
self.Window[row, col] += Energy # set the energy to the corresponding window
# stitch together the images from each process
if self.myid != 0:
pypar.send(self.Window, 0)
else:
self.Visual = np.zeros(shape = (self.WindowRows, self.WindowCols))
RowEnd = self.WindowRows/self.numproc
self.Visual[:RowEnd, :] = self.Window.copy()
for Pr in range(1,self.numproc):
RowStart = self.WindowRows/self.numproc * Pr
RowEnd = self.WindowRows/self.numproc * (Pr+1)
self.Visual[RowStart:RowEnd,:] = pypar.receive(Pr)
print "Elapsed time for process %d is %f" % (self.myid, time.time() - self.start)
def master(self):
numcompleted = 0
#--- start slaves distributing the first work slot
for i in range(0, min(self.numprocs-1, self.numworks)):
work = i
slave= i+1
pypar.send(work, destination=slave, tag=PYPAR_WORKTAG)
print '[MASTER ]: sent first work "%s" to node %d' %(work, slave)
# dispatch the remaining work slots on dynamic load-balancing policy
# the quicker to do the job, the more jobs it takes
for work in range(self.numprocs-1, self.numworks):
result, status = pypar.receive(source=pypar.any_source, tag=PYPAR_WORKTAG, return_status=True)
print '[MASTER ]: received result from node %d' %(status.source, )
#print result
numcompleted += 1
pypar.send(work, destination=status.source, tag=PYPAR_WORKTAG)
if self.debug: print '[MASTER ]: sent work "%s" to node %d' %(work, status.source)
self.work.handleWorkResult(result, status)
# all works have been dispatched out
print '[MASTER ]: ToDo : %d' %self.numworks
print '[MASTER ]: Done : %d' %numcompleted
# I've still to take into the remaining completions
while (numcompleted < self.numworks):
result, status = pypar.receive(source=pypar.any_source, tag=PYPAR_WORKTAG, return_status=True)
print '[MASTER ]: received (final) result from node %d' % (status.source, )
print result
numcompleted += 1
print '[MASTER ]: %d completed' %numcompleted
self.work.handleWorkResult(result, status)
print '[MASTER ]: about to terminate slaves'
# Tell slaves to stop working
for i in range(1, self.numprocs):
pypar.send('#', destination=i, tag=PYPAR_DIETAG)
if self.debug: print '[MASTER ]: sent DIETAG to node %d' %(i,)
def gather_dict(subdict, indexes):
"""
Recieve a dictionary from the children where the values are 1d arrays
and the arrays are chunks of the whole dictionary.
:param indexes: The indexes into the whole array.
:param subdict: The dictionary of 1d arrays to subset.
:returns: whole array
"""
if not STATE.is_parallel:
return subdict
else:
import pypar # pylint: disable=W0404
# Note, putting dictionary back sequentially
if STATE.rank == 0:
whole = {}
array_len = indexes[-1] # highest index in node 0 array
all_indexes = [[]] # Empty list for processor 0
for pro in range(1, STATE.size):
temp_indexes = pypar.receive(pro)
all_indexes.append(temp_indexes)
if temp_indexes[-1] > array_len:
array_len = temp_indexes[-1]
# Create the whole dictionary, filled with rank 0 info
for key in subdict.keys():
# Work-out the shape of arrays
array_shape = list(subdict[key].shape)
array_shape[0] = array_len + 1
whole[key] = numpy.empty(tuple(array_shape), subdict[key].dtype)
whole[key][indexes, ...] = subdict[key]
for pro in range(1, STATE.size):
subdict = pypar.receive(pro)
for key in whole.keys():
whole[key][all_indexes[pro], ...] = subdict[key]
return whole
else:
pypar.send(indexes, 0)
pypar.send(subdict, 0)
def gather_dict(subdict, indexes):
"""
Recieve a dictionary from the children where the values are 1d arrays
and the arrays are chunks of the whole dictionary.
:param indexes: The indexes into the whole array.
:param subdict: The dictionary of 1d arrays to subset.
:returns: whole array
"""
if not STATE.is_parallel:
return subdict
else:
import pypar # pylint: disable=W0404
# Note, putting dictionary back sequentially
if STATE.rank == 0:
whole = {}
array_len = indexes[-1] # highest index in node 0 array
all_indexes = [[]] # Empty list for processor 0
for pro in range(1, STATE.size):
temp_indexes = pypar.receive(pro)
all_indexes.append(temp_indexes)
if temp_indexes[-1] > array_len:
array_len = temp_indexes[-1]
# Create the whole dictionary, filled with rank 0 info
for key in list(subdict.keys()):
# Work-out the shape of arrays
array_shape = list(subdict[key].shape)
array_shape[0] = array_len + 1
whole[key] = numpy.empty(tuple(array_shape), subdict[key].dtype)
whole[key][indexes, ...] = subdict[key]
for pro in range(1, STATE.size):
subdict = pypar.receive(pro)
for key in list(whole.keys()):
whole[key][all_indexes[pro], ...] = subdict[key]
return whole
else:
pypar.send(indexes, 0)
pypar.send(subdict, 0)
def _get_statistic(self, source, statistic):
if self.is_master:
self.work_locally = True
statistic_by_station = statistician.DatabaseStatistician._get_statistic(self, source, statistic)
self.work_locally = False
for sender in range(1, pypar.size()):
statistic_by_station.update(pypar.receive(sender))
else:
if len(self.stations_local) > 0:
statistic_by_station = statistician.DatabaseStatistician._get_statistic(self, source, statistic)
else:
statistic_by_station = {}
pypar.send(statistic_by_station, root)
return statistic_by_station
def loadSkew(self, numberTasks): '''Computes the skew in the number of tasks to be processed by each node The skew is the standard deviation of the task number across the nodes''' if Environment.isParallel: import pypar if self.isRoot(): taskDistribution = [numberTasks] for i in range(1, pypar.size()): numberTasks = pypar.receive(i) taskDistribution.append(numberTasks) mean = reduce(operator.add, taskDistribution) mean = mean/float(len(taskDistribution)) #Std. dev stdev = 0 for el in taskDistribution: stdev = stdev + math.pow((el - mean), 2) skew = stdev/float(len(taskDistribution)) skew = math.sqrt(skew) for i in range(1, pypar.size()): pypar.send(skew, i) else: #Send the fragment pypar.send(numberTasks, 0) #Retrieve the array skew = pypar.receive(0) else: skew = 0 return skew
def combineDictionary(self, dictFragment): '''Combines a set of arrayFragments from each processor into one array''' if Environment.isParallel: import pypar if self.isRoot(): completeDict = dictFragment for i in range(1, pypar.size()): fragment = pypar.receive(i) completeDict.update(fragment) #Send the array for i in range(1, pypar.size()): pypar.send(completeDict, i) else: #Send the fragment pypar.send(dictFragment, 0) #Retrieve the array completeDict = pypar.receive(0) else: completeDict = dictFragment return completeDict
def grad_func(x0, *grad_args):
sys.stdout.flush()
x0 = numpy.asarray(x0)
# Generate a list of parameter sets to evaluate the function at
x_todo = [x0]
for ii in range(len(x0)):
eps = numpy.zeros(len(x0), numpy.float_)
eps[ii] = epsilon
x_todo.append(x0 + eps)
# Break that list apart into work for each node
x_by_node = []
for node_ii in range(num_procs):
x_by_node.append(x_todo[node_ii::num_procs])
# Master sends orders to all other nodes.
command = 'eval_func_over_list(func_str, x_todo, *grad_args)'
for node_ii in range(1, num_procs):
x_this_node = x_by_node[node_ii]
arguments = {'func_str': func_str,
'x_todo': x_this_node}
if send_grad_args:
arguments['grad_args'] = grad_args
pypar.send((command, arguments), node_ii)
sys.stdout.flush()
# This will hold the function evaluations done by each node.
vals_by_node = []
# The master node does its share of the work now
vals_by_node.append(eval_func_over_list(func_str, x_by_node[0],
*grad_args))
# Now receive the work done by each of the other nodes
for node_ii in range(1, num_procs):
vals_by_node.append(pypar.receive(node_ii))
# Reform the function value list that's broken apart by node.
func_evals = numpy.zeros(len(x0)+1, numpy.float_)
for node_ii,vals_this_node in enumerate(vals_by_node):
func_evals[node_ii::num_procs] = vals_this_node
# Now calculate the gradient
grad = numpy.zeros(len(x0), numpy.float_)
f0 = func_evals[0]
for ii,func_val in enumerate(func_evals[1:]):
grad[ii] = (func_val - f0)/epsilon
return grad
def collect_arr(arr):
"""
A useful collection routine for pypar.
If you are using pypar to parallelize the set of nested loops and fill
the resulting array, you usually need to combine the resulting array
from several mpi threads. In that case you just can execute
res=collect_arr(res)
And it will add the arrays from all the threads and store them in
the thread number 0
"""
if pypar.rank() > 0:
pypar.send(arr, 0)
else:
for i in range(1, pypar.size()):
arr = arr + pypar.receive(i)
return arr
def broadcast(self, data, process): '''Broadcasts data from process to all other nodes''' if Environment.isParallel: import pypar if self.rank() == process: #NOTE: pypar broadcast won't work even when setting vanilla #It always returns message trucated error. for i in range(pypar.size()): if i != self.rank(): pypar.send(data, i) else: data = pypar.receive(process) return data
def collect_arr(arr):
"""
A useful collection routine for pypar.
If you are using pypar to parallelize the set of nested loops and fill
the resulting array, you usually need to combine the resulting array
from several mpi threads. In that case you just can execute
res=collect_arr(res)
And it will add the arrays from all the threads and store them in
the thread number 0
"""
if pypar.rank() > 0:
pypar.send(arr, 0)
else:
for i in range(1, pypar.size()):
arr = arr + pypar.receive(i)
return arr
def grad_func(x0, *grad_args):
sys.stdout.flush()
x0 = numpy.asarray(x0)
# Generate a list of parameter sets to evaluate the function at
x_todo = [x0]
for ii in range(len(x0)):
eps = numpy.zeros(len(x0), numpy.float_)
eps[ii] = epsilon
x_todo.append(x0 + eps)
# Break that list apart into work for each node
x_by_node = []
for node_ii in range(num_procs):
x_by_node.append(x_todo[node_ii::num_procs])
# Master sends orders to all other nodes.
command = 'eval_func_over_list(func_str, x_todo, *grad_args)'
for node_ii in range(1, num_procs):
x_this_node = x_by_node[node_ii]
arguments = {'func_str': func_str, 'x_todo': x_this_node}
if send_grad_args:
arguments['grad_args'] = grad_args
pypar.send((command, arguments), node_ii)
sys.stdout.flush()
# This will hold the function evaluations done by each node.
vals_by_node = []
# The master node does its share of the work now
vals_by_node.append(
eval_func_over_list(func_str, x_by_node[0], *grad_args))
# Now receive the work done by each of the other nodes
for node_ii in range(1, num_procs):
vals_by_node.append(pypar.receive(node_ii))
# Reform the function value list that's broken apart by node.
func_evals = numpy.zeros(len(x0) + 1, numpy.float_)
for node_ii, vals_this_node in enumerate(vals_by_node):
func_evals[node_ii::num_procs] = vals_this_node
# Now calculate the gradient
grad = numpy.zeros(len(x0), numpy.float_)
f0 = func_evals[0]
for ii, func_val in enumerate(func_evals[1:]):
grad[ii] = (func_val - f0) / epsilon
return grad
def slave(self):
if self.debug: print '[SLAVE %d]: I am processor %d of %d on node %s' % (self.myid, self.myid, self.numprocs, self.node)
if self.debug: print '[SLAVE %d]: Entering work loop' % (self.myid,)
while True:
result, status = pypar.receive(source=0, tag=pypar.any_tag, return_status=True)
print '[SLAVE %d]: received work with tag %d from node %d'\
%(self.myid, status.tag, status.source)
if (status.tag == PYPAR_DIETAG):
print '[SLAVE %d]: received termination from node %d' % (self.myid, 0)
return
else:
worknum = result
if self.debug: print '[SLAVE %d]: work number is %s' % (self.myid, worknum)
myresult = self.work.calcWorkResult(worknum)
pypar.send(myresult, destination=0)
if self.debug: print '[SLAVE %d]: sent result to node %d' % (self.myid, 0)
def calc_num_blocks(self):
"""
pre-req: calc_lo_hi has been calculated - and only calculated once!
"""
if self.is_parallel is True:
import pypar
# print "synchronise self.rank", self.rank
if self.rank == 0:
calc_num_blocks = self._make_block_file
for source in range(1, self.size):
# print "waiting.."
received = pypar.receive(source)
# print "received", received
calc_num_blocks += received
return calc_num_blocks
else:
# print "sending from ", self.rank
pypar.send(self._make_block_file, 0)
def collected(iterable):
"""
Collect iterables back to the master.
"""
P = pp.size()
if P == 1:
for el in iterable:
yield el
else:
if pp.rank() == 0:
results = list(iterable)
for p in range(1, P):
pres = pp.receive(p)
results.extend(pres)
for el in results:
yield el
else:
pp.send(list(iterable), 0)
def calc_num_blocks(self):
"""
pre-req: calc_lo_hi has been calculated - and only calculated once!
"""
if self.is_parallel is True:
import pypar
# print "synchronise self.rank", self.rank
if self.rank == 0:
calc_num_blocks = self._make_block_file
for source in range(1, self.size):
# print "waiting.."
received = pypar.receive(source)
# print "received", received
calc_num_blocks += received
return calc_num_blocks
else:
# print "sending from ", self.rank
pypar.send(self._make_block_file, 0)
def ensemble_trajs(net, times, ensemble):
"""
Return a list of trajectories evaluated at times for all parameter sets
in ensemble.
"""
traj_set = []
elems_assigned = [ensemble[node::num_procs] for node in range(num_procs)]
for worker in range(1, num_procs):
command = 'Ensembles.few_ensemble_trajs(net, times, elements)'
args = {'net': net, 'times': times, 'elements': elems_assigned[worker]}
pypar.send((command, args), worker)
traj_set = few_ensemble_trajs(net, times, elems_assigned[0])
for worker in range(1, num_procs):
traj_set.extend(pypar.receive(worker))
return traj_set
