def __init__(self, verbose=True):
'''Initialises new Environment objects'''
self.outputDirectory = os.getcwd()
self.verbose = verbose
#Do the subsequent things once only
if not Environment.isInitialised:
#See if a parallel environment is available
try:
import pypar
#If its imported we might have a parallel environment
Environment.isParallel = True
self.output('[PEAT-SA] Parallel environment available')
#Check environment size.
#If there is more than one processor there must be a parallel environment
#If there's only one then its not parallel.
if pypar.size() == 1:
self.output('[PEAT-SA] Only one processor - parallel environment disabled')
Environment.isParallel = False
else:
self.output('[PEAT-SA] Parallel environment enabled with %d processors' % pypar.size())
except BaseException:
#Importing pypar caused an exception - No parallel environment
Environment.isParallel = False
self.output('[PEAT-SA] Parallel environment disabled.\n')
Environment.isInitialised = True
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 __init__(self, is_parallel=True):
"""
Use is_parallel = False to stop parallelism, eg when running
several scenarios.
"""
if is_parallel is True:
try:
import pypar
except ImportError:
self._not_parallel()
else:
if pypar.size() >= 2:
self.rank = pypar.rank()
self.size = pypar.size()
self.node = pypar.get_processor_name()
self.is_parallel = True
self.file_tag = FILE_TAG_DELIMITER + str(self.rank)
self.log_file_tag = FILE_TAG_DELIMITER + str(self.rank)
else:
self._not_parallel()
else:
self._not_parallel()
# Some constants to identify messages
self.load_event_set = 0
def _divideArray(self, array): '''Divides an array roughly equally depending on the environment size Returns: A list with one entry for each node. The entry is a tuple giving the start and end elements in array that should be assigned to that node.''' import pypar #Divide evenly then add remainder elements to processors maxElements = int(math.floor(len(array)/pypar.size())) remainder = len(array) - maxElements*pypar.size() start = 0 end = 0 divisions = [] for i in range(pypar.size()): start = end end = end + maxElements if remainder != 0: end = end + 1 remainder = remainder - 1 divisions.append((start, end)) return divisions
def __init__(self, is_parallel=True):
"""
Use is_parallel = False to stop parallelism, eg when running
several scenarios.
"""
if is_parallel is True:
try:
import pypar
except ImportError:
self._not_parallel()
else:
if pypar.size() >= 2:
self.rank = pypar.rank()
self.size = pypar.size()
self.node = pypar.get_processor_name()
self.is_parallel = True
self.file_tag = FILE_TAG_DELIMITER + str(self.rank)
self.log_file_tag = FILE_TAG_DELIMITER + str(self.rank)
else:
self._not_parallel()
else:
self._not_parallel()
# Some constants to identify messages
self.load_event_set = 0
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 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 _mpi_end_embarrass():
global _mpi_initialized
if _mpi_initialized:
import pypar
print(pypar.rank() + 1, " of ", pypar.size(), ": BARRIER")
pypar.barrier()
print(pypar.rank() + 1, " of ", pypar.size(), ": FINALIZE")
pypar.finalize()
_mpi_initialized = False
else:
print("Non-MPI run : Exit without MPI_Finalize")
def __init__(self,
coordinates,
vertices,
boundary = None,
full_send_dict = None,
ghost_recv_dict = None,
velocity = None):
Domain.__init__(self,
coordinates,
vertices,
boundary,
velocity = velocity,
full_send_dict=full_send_dict,
ghost_recv_dict=ghost_recv_dict,
processor=pypar.rank(),
numproc=pypar.size()
)
N = self.number_of_elements
self.communication_time = 0.0
self.communication_reduce_time = 0.0
print 'processor',self.processor
print 'numproc',self.numproc
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 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 __init__(self, LookPos, LookDir, LookYaw, WindowRows = 40, WindowCols = 40):
self.LookPos = np.array(LookPos)
self.LookDir = np.array(LookDir)
self.Yaw = LookYaw
self.WindowRows = WindowRows
self.WindowCols = WindowCols
rhop = np.linalg.norm(np.array([LookDir[0],LookDir[1]]))
self.__Lon = math.atan2(LookDir[1], LookDir[0])
self.__Lat = math.atan2(LookDir[2],rhop)
self.start = time.time()
# initialize the MPI
self.numproc = pypar.size()
self.myid = pypar.rank()
self.node = pypar.get_processor_name()
if self.myid != self.numproc - 1:
self.Rows = self.WindowRows/self.numproc
self.RowEnd = self.WindowRows/self.numproc * (self.myid+1) - 1
else:
self.Rows = self.WindowRows/self.numproc + self.WindowRows%self.numproc
self.RowEnd = self.WindowRows
self.RowStart = self.WindowRows/self.numproc * self.myid
self.Window = np.zeros(shape = (self.Rows, self.WindowCols))
def rec_submesh(p, verbose=True):
import pypar
numproc = pypar.size()
myid = pypar.rank()
[submesh_cell, triangles_per_proc,\
number_of_full_nodes, number_of_full_triangles] = rec_submesh_flat(p,verbose)
# find the full triangles assigned to this processor
lower_t = 0
for i in range(myid):
lower_t = lower_t+triangles_per_proc[i]
upper_t = lower_t+triangles_per_proc[myid]
# convert the information into a form needed by the GA
# datastructure
[GAnodes, GAtriangles, boundary, quantities, \
ghost_rec, full_send, \
tri_map, node_map, tri_l2g, node_l2g, \
ghost_layer_width] = \
build_local_mesh(submesh_cell, lower_t, upper_t, numproc)
return GAnodes, GAtriangles, boundary, quantities,\
ghost_rec, full_send,\
number_of_full_nodes, number_of_full_triangles, tri_map, node_map,\
tri_l2g, node_l2g, ghost_layer_width
def __init__(self,
coordinates,
vertices,
boundary=None,
full_send_dict=None,
ghost_recv_dict=None,
velocity=None):
Domain.__init__(self,
coordinates,
vertices,
boundary,
velocity=velocity,
full_send_dict=full_send_dict,
ghost_recv_dict=ghost_recv_dict,
processor=pypar.rank(),
numproc=pypar.size())
N = self.number_of_elements
self.communication_time = 0.0
self.communication_reduce_time = 0.0
print 'processor', self.processor
print 'numproc', self.numproc
def rec_submesh(p, verbose=True):
import pypar
numproc = pypar.size()
myid = pypar.rank()
[submesh_cell, triangles_per_proc,\
number_of_full_nodes, number_of_full_triangles] = rec_submesh_flat(p,verbose)
# find the full triangles assigned to this processor
lower_t = 0
for i in range(myid):
lower_t = lower_t + triangles_per_proc[i]
upper_t = lower_t + triangles_per_proc[myid]
# convert the information into a form needed by the GA
# datastructure
[GAnodes, GAtriangles, boundary, quantities, \
ghost_rec, full_send, \
tri_map, node_map, tri_l2g, node_l2g, \
ghost_layer_width] = \
build_local_mesh(submesh_cell, lower_t, upper_t, numproc)
return GAnodes, GAtriangles, boundary, quantities,\
ghost_rec, full_send,\
number_of_full_nodes, number_of_full_triangles, tri_map, node_map,\
tri_l2g, node_l2g, ghost_layer_width
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 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 broadcast( obj ): if root(): for i in xrange(p.size()-1): p.send(obj,i+1) return obj else: return p.receive(0)
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 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 _get_stations_local(self):
all_stations = self._get_all_stations()
num_stations = len(all_stations)
mpisize = pypar.size()
stations_per_proc = num_stations / mpisize
leftover = num_stations % mpisize
_station_distr = []
ind_start = 0
ind_end = 0
rank = pypar.rank()
for ind_proc in range(mpisize):
ind_start = ind_end
count = stations_per_proc
if ind_proc < leftover:
count += 1
ind_end = ind_start + count
_station_distr.append((ind_start, ind_end))
start_local, end_local = _station_distr[rank]
print pypar.rank(), stations_per_proc, leftover, start_local, end_local
if start_local >= num_stations:
self.stations_local = set()
self.say('No enough stations for this process')
else:
self.say('Station range: %i -- %i (%i)' % (start_local, end_local, num_stations))
self.stations_local = set(all_stations[start_local:end_local])
assert self.stations_local or not self.is_master
def __init__(self):
self.proc = pypar.size()
self.myid = pypar.rank()
self.node = pypar.get_processor_name()
return
def run():
"""
Run the process, handling any parallelisation.
"""
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("-c", "--config",
help="Configuration file",
type=str)
parser.add_argument("-i", "--inputfile",
help="Input DEM file (ascii format)",
type=str)
parser.add_argument("-o", "--output",
help="Output path",
type=str)
parser.add_argument("-v", "--verbose",
help=("Verbose output (not available when invoking"
"parallel run)") )
args = parser.parse_args()
logfile = 'topomult.log'
loglevel = 'INFO'
if args.verbose:
verbose = args.verbose
else:
verbose = False
if args.config:
cfg = ConfigParser.ConfigParser()
cfg.read(args.config)
input_file = cfg.get('Input', 'Filename')
output_path = cfg.get('Output', 'Path')
logfile = cfg.get('Logging', 'LogFile')
loglevel = cfg.get('Logging', 'LogLevel')
verbose = cfg.get('Logging', 'Verbose')
if args.inputfile:
input_file = args.inputfile
if args.output:
output_path = args.output
attemptParallel()
if pp.size() > 1 and pp.rank() > 0:
logfile += '-' + str(pp.rank())
verbose = False # to stop output to console
flStartLog(logfile, loglevel, verbose)
pp.barrier()
work(input_file, output_path,
['n','s','e','w','ne','nw','se','sw'])
pp.barrier()
pp.finalize()
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 all_gather(obj):
myid = p.rank()
nproc = p.size()
result = [None for i in xrange(nproc)]
result[myid] = obj
for i in xrange(nproc):
broadcast_vec(result, i)
return result
def all_gather( obj ): myid = p.rank() nproc = p.size() result = [ None for i in xrange(nproc) ] result[myid] = obj for i in xrange(nproc): broadcast_vec(result,i) return result
def Inlet_operator(domain,
poly,
Q,
velocity=None,
default=None,
description=None,
label=None,
logging=False,
master_proc=0,
procs=None,
verbose=False):
# If not parallel domain then allocate serial Inlet operator
if isinstance(domain, Parallel_domain) is False:
if verbose: print "Allocating non parallel inlet operator ....."
return anuga.structures.inlet_operator.Inlet_operator(
domain,
poly,
Q,
velocity=velocity,
default=default,
description=description,
label=label,
logging=logging,
verbose=verbose)
import pypar
if procs is None:
procs = range(0, pypar.size())
myid = pypar.rank()
poly = num.array(poly, dtype='d')
alloc, inlet_master_proc, inlet_procs, enquiry_proc = allocate_inlet_procs(
domain, poly, master_proc=master_proc, procs=procs, verbose=verbose)
if alloc:
if verbose and myid == inlet_master_proc:
print "Parallel Inlet Operator ================="
print "Poly = " + str(poly)
print "Master Processor is P%d" % (inlet_master_proc)
print "Processors are P%s" % (inlet_procs)
print "========================================="
return Parallel_Inlet_operator(domain,
poly,
Q,
velocity=velocity,
default=default,
description=description,
label=label,
logging=logging,
master_proc=inlet_master_proc,
procs=inlet_procs,
verbose=verbose)
else:
return None
def run_client():
'''
Runs
'''
# Identification
myid = pypar.rank() # id of this process
nproc = pypar.size() # number of processors
print "I am client", myid
pypar.finalize()
def run_client():
'''
Runs
'''
# Identification
myid = pypar.rank() # id of this process
nproc = pypar.size() # number of processors
print "I am client", myid
pypar.finalize()
def p_sum_all(vec):
nproc = p.size()
pcs = [None for i in xrange(nproc)]
if root():
for i in xrange(nproc):
pcs[i] = vec[i::nproc]
temp = scatter(pcs)
temp = sum(temp)
pcs = all_gather(temp)
return sum(pcs)
def calc_participant_list(s):
for rank in xrange(1, mpi.size()):
N1 = s.Nx_sum_list[rank - 1]
N2 = s.Nx_sum_list[rank]
if s.gi1 >= N1 and s.gi1 <= N2:
start_rank = rank
if s.gi2 >= N1 and s.gi2 <= N2:
end_rank = rank
return range(start_rank, end_rank + 1)
def one_example():
txt = ["yes", "no", "when", "what the", "a", "5ive!"]
rank = pypar.rank()
size = pypar.size()
print "I am processor %d of %d. " % (rank, size)
for i, ele in enumerate(txt):
if i % size == rank:
print "i" + str(i) + " P" + str(rank) + " len " + str(len(ele)) + " for " + ele
def calc_participant_list( s ): for rank in xrange( 1, mpi.size() ): N1 = s.Nx_sum_list[rank-1] N2 = s.Nx_sum_list[rank] if s.gi1 >= N1 and s.gi1 <= N2: start_rank = rank if s.gi2 >= N1 and s.gi2 <= N2: end_rank = rank return range( start_rank, end_rank+1 )
def p_sum_all( vec ): nproc = p.size() pcs = [ None for i in xrange(nproc) ] if root(): for i in xrange(nproc): pcs[i] = vec[i::nproc] temp = scatter(pcs) temp = sum(temp) pcs = all_gather(temp) return sum(pcs)
def one_example():
txt = ["yes", "no", "when", "what the", "a", "5ive!"]
rank = pypar.rank()
size = pypar.size()
print "I am processor %d of %d. " % (rank, size)
for i, ele in enumerate(txt):
if i % size == rank:
print "i" + str(i) + " P" + str(rank) + " len " + str(
len(ele)) + " for " + ele
def test_string(self):
myid, ncpu = pp.rank(), pp.size()
data = 'ABCDEFGHIJKLMNOP' # Length = 16
NP = len(data) / ncpu
X = ' ' * NP
pp.scatter(data, 0, buffer=X) # With buffer
Y = pp.scatter(data, 0) # With buffer automatically created
self.assertEqual(X, Y)
self.assertEqual(Y, data[myid * NP:(myid + 1) * NP])
self.assertEqual(X, data[myid * NP:(myid + 1) * NP])
def test_string(self):
myid, ncpu = pp.rank(), pp.size()
data = 'ABCDEFGHIJKLMNOP' # Length = 16
NP = len(data) / ncpu
X = ' ' * NP
pp.scatter(data, 0, buffer=X) # With buffer
Y = pp.scatter(data, 0) # With buffer automatically created
self.assertEqual(X, Y)
self.assertEqual(Y, data[myid * NP:(myid + 1) * NP])
self.assertEqual(X, data[myid * NP:(myid + 1) * NP])
def test_without_root(self):
myid, ncpu = pp.rank(), pp.size()
N = 16
NP = N / ncpu
data = np.array(range(N)).astype('i')
X = np.zeros(NP).astype('i')
pp.scatter(data, buffer=X) # With buffer
Y = pp.scatter(data) # With buffer automatically created
self.assertTrue(np.allclose(X, Y))
self.assertTrue(np.allclose(X, data[myid * NP:(myid + 1) * NP]))
self.assertTrue(np.allclose(Y, data[myid * NP:(myid + 1) * NP]))
def test_without_root(self):
myid, ncpu = pp.rank(), pp.size()
N = 16
NP = N / ncpu
data = np.array(range(N)).astype('i')
X = np.zeros(NP).astype('i')
pp.scatter(data, buffer=X) # With buffer
Y = pp.scatter(data) # With buffer automatically created
self.assertTrue(np.allclose(X, Y))
self.assertTrue(np.allclose(X, data[myid * NP:(myid + 1) * NP]))
self.assertTrue(np.allclose(Y, data[myid * NP:(myid + 1) * NP]))
def p_dot_all( a, b ): nproc = p.size() va = [ None for i in xrange(nproc) ] vb = [ None for i in xrange(nproc) ] if root(): for i in xrange(nproc): va[i] = a[i::nproc] vb[i] = b[i::nproc] ta = scatter(va) tb = scatter(vb) pv = [ ta[i]*tb[i] for i in xrange(len(ta)) ] ps = sum(pv) rv = all_gather(ps) return sum(rv)
def p_dot_all(a, b):
nproc = p.size()
va = [None for i in xrange(nproc)]
vb = [None for i in xrange(nproc)]
if root():
for i in xrange(nproc):
va[i] = a[i::nproc]
vb[i] = b[i::nproc]
ta = scatter(va)
tb = scatter(vb)
pv = [ta[i] * tb[i] for i in xrange(len(ta))]
ps = sum(pv)
rv = all_gather(ps)
return sum(rv)
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 __init__(self):
"""
Use is_parallel = False to stop parallelism, eg when running
several scenarios.
"""
try:
import pypar # pylint: disable=W0404
except ImportError:
self._not_parallel()
else:
if pypar.size() >= 2:
self.rank = pypar.rank()
self.size = pypar.size()
self.node = pypar.get_processor_name()
self.is_parallel = True
self.file_tag = str(self.rank)
self.log_file_tag = str(self.rank)
# Ensure a clean MPI exit
atexit.register(pypar.finalize)
else:
self._not_parallel()
def __init__(self):
"""
Use is_parallel = False to stop parallelism, eg when running
several scenarios.
"""
try:
import pypar # pylint: disable=W0404
except ImportError:
self._not_parallel()
else:
if pypar.size() >= 2:
self.rank = pypar.rank()
self.size = pypar.size()
self.node = pypar.get_processor_name()
self.is_parallel = True
self.file_tag = str(self.rank)
self.log_file_tag = str(self.rank)
# Ensure a clean MPI exit
atexit.register(pypar.finalize)
else:
self._not_parallel()
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 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 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 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 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 allreduce(x, op, buffer=None, vanilla=0, bypass=False):
"""Allreduce elements in x to buffer (of the same size as x)
applying operation op elementwise.
If bypass is True, all admin and error checks
get bypassed to reduce the latency.
The buffer must be specified explicitly in this case.
"""
if bypass:
allreduce_array(x, buffer, op)
return
import types
from pypar import size
numproc = size() # Needed to determine buffer size
# Create metadata about object
protocol, typecode, size, shape = create_control_info(x)
# Allreduce
if protocol == 'array':
if buffer is None:
buffer = zeros(size*numproc, typecode)
# Modify shape along axis=0 to match size
shape = list(shape)
shape[0] *= numproc
buffer = reshape(buffer, shape)
msg = 'Data array and buffer must have same type '
msg = 'in allreduce. I got types "%s" and "%s"' % (x.dtype.char,
buffer.dtype.char)
assert x.dtype.char == buffer.dtype.char, msg
allreduce_array(x, buffer, op)
elif (protocol == 'vanilla' or protocol == 'string'):
raise 'Protocol: %s unsupported for allreduce' % protocol
else:
raise 'Unknown protocol: %s' % protocol
return buffer
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 test_diff_master(self):
myid, ncpu = pp.rank(), pp.size()
N = 16
NP = N / ncpu
check = np.array(range(N)).astype('i')
X = np.zeros(NP).astype('i')
data = np.empty(N, 'i')
if myid == 0: # only generated on master
data = np.array(range(N)).astype('i')
pp.scatter(data, 0, buffer=X) # With buffer
Y = pp.scatter(data, 0) # With buffer automatically created
self.assertTrue(np.allclose(X, Y))
self.assertTrue(np.allclose(X, check[myid * NP:(myid + 1) * NP]))
self.assertTrue(np.allclose(Y, check[myid * NP:(myid + 1) * NP]))
def embarrass(file_list, pre_roll=0, post_roll=0):
"""Get a file list for this node (embarrassing parallelization)"""
global _mpi_initialized
if type(file_list) == set:
file_list = list(file_list)
import sys
try:
proc = int(sys.argv[-2])
of = int(sys.argv[-1])
except (IndexError, ValueError):
if pre_roll != 0 or post_roll != 0:
raise AssertionError(
"Pre/post-roll no longer supported for MPI -- non-contiguuous")
print("Trying to run in MPI mode...")
import pypar
_mpi_initialized = True
proc = pypar.rank() + 1
of = pypar.size()
print("Success!", proc, "of", of)
return _mpi_iterate(file_list)
i = (len(file_list) * (proc - 1)) / of
j = (len(file_list) * proc) / of - 1
assert proc <= of and proc > 0
if proc == of:
j += 1
print(proc, "processing", i, j, "(inclusive)")
i -= pre_roll
j += post_roll
if i < 0:
i = 0
if j >= len(file_list):
j = len(file_list) - 1
return file_list[i:j + 1]