def execTask(self, task):
"""Wrapper function calling mapping/reducing/finalizing phase tasks,
dispatch tasks to workers until all finished and collect feedback.
Faulty workers are removed from active duty work list.
"""
atimer = Timer(task)
print( 'Entering {0:s} phase...'.format(task) )
taskDict = { 'Map':(self.mapIn, MAP_START, MAP_FINISH), \
'Init':(self.mapIn, INIT_START, MAP_FINISH), \
'Reduce':(self.reduceIn, REDUCE_START, REDUCE_FINISH) }
# line up jobs and workers into priority queues
jobs = taskDict[task][0][:]
heapq.heapify(jobs); running = {}
heapq.heapify(self.workers)
while (jobs or running) and self.nActive > 0:
# dispatch all jobs to all free workers
while jobs and self.workers[0].isFree():
job = heapq.heappop(jobs)
worker = heapq.heappop(self.workers)
world.send(job, dest=worker.id, tag=taskDict[task][1])
worker.setBusy(); heapq.heappush(self.workers, worker)
running[job] = (time(), worker)
if self.config['verbosity'] >= 6:
print('Dispatching file '+os.path.basename(job)+' to worker '+str(worker.id))
# if no more free workers, break
if not self.workers[0].isFree(): break
# wait for finishing workers as well as do cleaning
self.wait(running, taskDict[task][2])
self.clean(running, jobs)
print( '{0:s} phase completed'.format(task) )
def map(self, tag):
"""
Execute supplied mapfn on each key-value pair read from file
assigned by the master node
"""
atimer = Timer('Worker_Map')
# load key-value pairs from filename
filename = world.recv(source=0, tag=tag)
data = self.read(filename, tag)
buffer = [ [] for ii in range(self.config['nReduce']) ]
for key, val in data.items():
for newKey, newVal in self.config['mapfn'](key, val):
idx = self.config['hashfn'](newKey) % self.config['nReduce']
buffer[idx].append( (newKey, newVal) )
# write out new key-value pairs in scattered files
for ii in range(self.config['nReduce']):
tmpfile = self.reduceIn[ii]+'-tmp'+str(world.rank)
# dump in append mode
with open(tmpfile, 'a+b') as fout:
pickle.dump(buffer[ii], fout, pickle.HIGHEST_PROTOCOL)
# report back as successful completion of task
world.send(filename, dest=0, tag=MAP_FINISH)
def map(self, f, tasks):
N = len(tasks)
P = self.P
Pless1 = P - 1
if self.rank != 0:
self.wait()
return
if f is not self.f:
self.f = f
requests = []
for p in range(1, self.P):
r = COMM_WORLD.isend(f, dest=p)
requests.append(r)
MPI.Request.waitall(requests)
requests = []
for i, task in enumerate(tasks):
r = COMM_WORLD.isend(task, dest=(i%Pless1)+1, tag=i)
requests.append(r)
MPI.Request.waitall(requests)
results = []
for i in range(N):
result = COMM_WORLD.recv(source=(i%Pless1)+1, tag=i)
results.append(result)
return results
def test_padded_kernel(self):
""" Implement a simple padded kernel. """
for case in self.cases:
# Form data to work on.
space.initialize_space(case['shape'])
x_np = comm.allreduce(np.random.randn(*case['shape']).astype(case['dtype']))
x = Grid(x_np, x_overlap=1)
s_np = comm.allreduce(np.random.randn(1).astype(case['dtype']))
s = Const(s_np)
z = Out(case['dtype'])
# Make a kernel.
code = Template("""
if (_in_local && _in_global) {
x(0,0,0) = s(0) * x(0,0,0);
z += a * x(0,0,0);
}
""").render()
fun = Kernel(code, \
('a', 'number', case['dtype']), \
('x', 'grid', x.dtype), \
('s', 'const', s.dtype, s.data.size), \
('z', 'out', z.dtype), \
padding=(1,1,1,1))
# Execute and check the result.
fun(case['dtype'](2), x, s, z)
gpu_sum = z.get()
cpu_sum = np.sum(2.0 * s_np * x_np)
err = abs(gpu_sum - cpu_sum) / abs(cpu_sum)
# print case, err
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-2, (case, err))
else:
self.assertTrue(err < 1e-6, (case, err))
def sync(self, key):
"""Synchronize dataType with all worker nodes.
"""
tagDict = { 'Map': (UPDATE_MAP, self.mapIn),\
'Reduce': (UPDATE_REDUCE, self.reduceIn),\
'Config': (UPDATE_CONFIG, self.config) }
for ii in range(1, len(self.workers)+1):
world.send(tagDict[key][1], dest=ii, tag=tagDict[key][0])
def wait(self):
if self.rank == 0:
raise RuntimeError("Proc 0 cannot wait!")
status = MPI.Status()
while True:
task = COMM_WORLD.recv(source=0, tag=MPI.ANY_TAG, status=status)
if not task:
break
if isinstance(task, FunctionType):
self.f = task
continue
result = self.f(task)
COMM_WORLD.isend(result, dest=0, tag=status.tag)
def update(self, tag):
"""Update file list and global configurations
"""
atimer = Timer('Worker_Update')
if tag == UPDATE_MAP:
self.mapIn = world.recv(source=0, tag=tag)
elif tag == UPDATE_REDUCE:
self.reduceIn = world.recv(source=0, tag=tag)
self.reduceOut = [ os.path.splitext(file)[0]+'.red' for file in self.reduceIn ]
elif tag == UPDATE_CONFIG:
self.config = world.recv(source=0, tag=tag)
else:
raise ValueError('Wrong tag specified.')
def get_cpu_raw(cpu_data, k):
# Make sure overlapped data is accurate as well.
xr = space.get_space_info()['x_range']
if comm.Get_rank() == 0:
pad_back = cpu_data[-k:, :, :]
else:
pad_back = cpu_data[xr[0] - k:xr[0], :, :]
if comm.Get_rank() == comm.Get_size() - 1:
pad_front = cpu_data[:k, :, :]
else:
pad_front = cpu_data[xr[1]:xr[1] + k, :, :]
return np.concatenate((pad_back, cpu_data[xr[0]:xr[1],:,:], \
pad_front), axis=0)
def wait(self, running, tag):
"""Test if any worker has finished its job.
If so, decrease its key and make it available
"""
atimer = Timer('Wait')
inittime = time()
status = MPI.Status()
while time() - inittime < self.config['jobwait']:
if world.Iprobe(source=MPI.ANY_SOURCE,tag=tag,status=status):
jobf = world.recv(source=status.source, tag=tag)
idx = 0
for ii, worker in enumerate(self.workers):
if worker.id == status.source: idx = ii; break
if self.config['verbosity'] >= 8:
print('Freeing worker '+str(self.workers[idx].id))
worker = self.workers[idx]
# faulty worker's job has already been cleaned
if not worker.isFaulty():
del running[jobf]
else:
self.nActive += 1
worker.setFree()
heapq._siftup(self.workers, idx)
def _init_gpu(comm):
""" Chooses a gpu and creates a context on it. """
# Find out how many GPUs are available to us on this node.
driver.init()
num_gpus = driver.Device.count()
# Figure out the names of the other hosts.
rank = comm.Get_rank() # Find out which process I am.
name = MPI.Get_processor_name() # The name of my node.
hosts = comm.allgather(name) # Get the names of all the other hosts
# Find out which GPU to take (by precedence).
gpu_id = hosts[0:rank].count(name)
if gpu_id >= num_gpus:
raise TypeError("No GPU available.")
# Create a context on the appropriate device.
for k in range(num_gpus):
try:
device = driver.Device((gpu_id + k) % num_gpus)
context = device.make_context()
except:
continue
else:
# print "On %s: process %d taking gpu %d of %d.\n" % \
# (name, rank, gpu_id+k, num_gpus)
break
return device, context # Return device and context.
def test_batch_sum(self):
""" Make sure batch summing works. """
num_outs = 3
for case in self.cases:
space.initialize_space(case['shape'])
x = [Out(case['dtype'], op='sum') for k in range(num_outs)]
x_cpu_data = [np.random.randn(*case['shape'][1:])\
.astype(case['dtype']) for k in range(num_outs)]
if case['dtype'] in (np.complex64, np.complex128):
for k in range(num_outs):
x_cpu_data[k] = (1 + 1j) * x_cpu_data[k]
res_gold = []
for k in range(num_outs):
x[k].data.set(x_cpu_data[k])
res_gold.append(comm.allreduce(np.sum(x_cpu_data[k].flatten())))
batch_reduce(*x)
res_gpu = [x_indiv.get() for x_indiv in x]
for k in range(num_outs):
err = abs(res_gold[k] - res_gpu[k]) / abs(res_gold[k])
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-3)
else:
self.assertTrue(err < 1e-10)
def update(self, tag):
"""Update file list and global configurations
"""
atimer = Timer('Worker_Update')
if tag == UPDATE_MAP:
self.mapIn = world.recv(source=0, tag=tag)
elif tag == UPDATE_REDUCE:
self.reduceIn = world.recv(source=0, tag=tag)
self.reduceOut = [
os.path.splitext(file)[0] + '.red' for file in self.reduceIn
]
elif tag == UPDATE_CONFIG:
self.config = world.recv(source=0, tag=tag)
else:
raise ValueError('Wrong tag specified.')
def get(self):
""" Redefined so that we don't get overlap data. """
# Get our section of the grid (excluding overlap).
if self._xlap is 0:
data = self.data.get()
else:
data = self.data.get()[self._xlap:-self._xlap,:,:]
# return np.concatenate(comm.allgather(data), axis=0) # Super-simple.
result = comm.gather(data) # Gather all peices to root.
if comm.Get_rank() == 0:
# Root node glues everything together.
return np.concatenate(result, axis=0)
else:
return None
def debug(*s):
import sys
from mpi4py.MPI import COMM_WORLD
print('[rank:{}]'.format(COMM_WORLD.Get_rank()),
*s,
file=sys.stderr,
flush=True)
def run(self):
"""Receiving job instructions from the master node until
TERMINATE signal received. Allowed tasks are defined in taskDict
"""
atimer = Timer('Worker')
# tasks define signal-behavior in the run function
taskDict = { MAP_START: self.map, REDUCE_START: self.reduce,\
INIT_START: self.map,\
UPDATE_MAP: self.update, UPDATE_REDUCE: self.update,\
UPDATE_CONFIG: self.update }
status = MPI.Status()
while True:
# ping input
if not world.Iprobe(source=0, tag=MPI.ANY_TAG, status=status):
sleep(self.config['delay']);
# entire calculation finished
elif status.tag == TERMINATE:
term = world.recv(source=0, tag=TERMINATE); break
# check allowed tasks
elif status.tag in taskDict:
taskDict[status.tag](status.tag);
# no instruction found, looping
else:
sleep(self.config['delay'])
def main():
args = parse_args()
assert args.pretrained_model_path is None or args.pretrained_model_path.endswith(
".ckpt")
os.makedirs(args.save_dir, exist_ok=True)
save_args(args, args.save_dir)
set_seed(args.seed + COMM_WORLD.Get_rank() * 100)
nprocs = COMM_WORLD.Get_size()
# Initialize model and agent policy
aurora = Aurora(args.seed + COMM_WORLD.Get_rank() * 100, args.save_dir,
int(7200 / nprocs), args.pretrained_model_path,
tensorboard_log=args.tensorboard_log)
# training_traces, validation_traces,
training_traces = []
val_traces = []
if args.train_trace_file:
with open(args.train_trace_file, 'r') as f:
for line in f:
line = line.strip()
training_traces.append(Trace.load_from_file(line))
if args.val_trace_file:
with open(args.val_trace_file, 'r') as f:
for line in f:
line = line.strip()
if args.dataset == 'pantheon':
queue = 100 # dummy value
# if "ethernet" in line:
# queue = 500
# elif "cellular" in line:
# queue = 50
# else:
# queue = 100
val_traces.append(Trace.load_from_pantheon_file(
line, queue=queue, loss=0))
elif args.dataset == 'synthetic':
val_traces.append(Trace.load_from_file(line))
else:
raise ValueError
aurora.train(args.randomization_range_file,
args.total_timesteps, tot_trace_cnt=args.total_trace_count,
tb_log_name=args.exp_name, validation_flag=args.validation,
training_traces=training_traces,
validation_traces=val_traces,
real_trace_prob=args.real_trace_prob)
def main(argv=None):
args = process_command_line(argv)
# note that in MPI mode, lengths will be global, whereas data will
# be local (i.e. only this node's data).
lengths, data = load_trjs_or_features(args)
kwargs = {}
if args.cluster_iterations is not None:
kwargs['kmedoids_updates'] = int(args.cluster_iterations)
clustering = args.Clusterer(metric=args.cluster_distance,
n_clusters=args.cluster_number,
cluster_radius=args.cluster_radius,
mpi_mode=mpi_mode,
**kwargs)
clustering.fit(data)
# release the RAM held by the trajectories (we don't need it anymore)
del data
logger.info("Clustered %s frames into %s clusters in %s seconds.",
sum(lengths), len(clustering.centers_), clustering.runtime_)
result = clustering.result_
if mpi_mode:
local_ctr_inds, local_dists, local_assigs = \
result.center_indices, result.distances, result.assignments
with timed("Reassembled dist and assign arrays in %.2f sec",
logging.info):
all_dists = mpi.ops.assemble_striped_ragged_array(
local_dists, lengths)
all_assigs = mpi.ops.assemble_striped_ragged_array(
local_assigs, lengths)
ctr_inds = mpi.ops.convert_local_indices(local_ctr_inds, lengths)
result = ClusterResult(center_indices=ctr_inds,
distances=all_dists,
assignments=all_assigs,
centers=result.centers)
result = result.partition(lengths)
if mpi.rank() == 0:
with timed("Wrote center indices in %.2f sec.", logger.info):
write_centers_indices(args.center_indices,
[(t, f * args.subsample)
for t, f in result.center_indices])
with timed("Wrote center structures in %.2f sec.", logger.info):
write_centers(result, args)
write_assignments_and_distances_with_reassign(result, args)
mpi.comm.barrier()
logger.info("Success! Data can be found in %s.",
os.path.dirname(args.distances))
return 0
def RandomSize(N_lower, N_upper):
if COMM_WORLD.rank == 0:
size = randint(N_lower, N_upper)
else:
size = None
size = COMM_WORLD.bcast(size, root=0)
assert size is not None
return size
def RandomNumber():
if COMM_WORLD.rank == 0:
number = _rand(1)[0]
else:
number = None
number = COMM_WORLD.bcast(number, root=0)
assert number is not None
return number
def RandomTuple(Q):
if COMM_WORLD.rank == 0:
tuple_ = tuple(float(v) for v in _rand(Q))
else:
tuple_ = None
tuple_ = COMM_WORLD.bcast(tuple_, root=0)
assert tuple_ is not None
return tuple_
def __init__(self):
"""Obtain configurations and all filelists from the master node
"""
assert world.rank >= 1
# synchronize configuration file and file lists with the master node
self.config = {}; self.mapIn = []; self.reduceIn = []; self.reduceOut = []
self.config = world.bcast(self.config, root=0)
def simulate(name, check_success_only=False):
""" Read simulation from input file, simulate, and write out results. """
# Reset the environment variables pointing to the temporary directory.
tempfile.tempdir = '/tmp'
# Create the reporter function.
write_status = lambda msg: open(name + '.status', 'a').write(msg)
if comm.Get_rank() == 0:
# write_status('EXEC initializing\n')
def rep(err):
write_status('%e\n' % err)
else: # No reporting needed for non-root nodes.
def rep(err):
pass
# Get input parameters.
params = get_parameters(name)
# Define operations needed for the lumped bicg operation.
b, x, ops, post_cond = maxwell_ops_lumped.ops(params)
# Solve!
start_time = time.time()
rep.stime = start_time
x, err, success = bicg.solve_symm_lumped(b, x=x, \
max_iters=params['max_iters'], \
reporter=rep, \
err_thresh=params['err_thresh'], \
**ops)
if check_success_only: # Don't write output, just see if we got a success.
return success
# Gather results onto root's host memory.
result = { 'E': [E.get() for E in x], \
'err': err, \
'success': success}
# Write results to output file.
if comm.Get_rank() == 0:
result['E'] = post_cond(result['E']) # Apply postconditioner.
write_results(name, result)
return success
def array(self, N=0, filename=None, component=None, root=0):
"""Dump data to numpy format on root processor."""
assert(N == 0 or N == 1)
is_root = comm.Get_rank() == root
size = self.get_total_number_probes() if is_root else len(self)
comp = self.value_size() if component is None else 1
z = zeros((size, comp))
# Retrieve all values
if len(self) > 0:
for k in range(comp):
if is_root:
ids = self.get_probe_ids()
z[ids, k] = self.get_probes_component_and_snapshot(k, N)
else:
z[:, k] = self.get_probes_component_and_snapshot(k, N)
# Collect on root
recvfrom = comm.gather(len(self), root=root)
if is_root:
for j, k in enumerate(recvfrom):
if comm.Get_rank() != j:
ids = comm.recv(source=j, tag=101)
z0 = comm.recv(source=j, tag=102)
z[ids, :] = z0[:, :]
else:
ids = self.get_probe_ids()
comm.send(ids, dest=root, tag=101)
comm.send(z, dest=root, tag=102)
if is_root:
if filename:
z.dump(filename+"_statistics.probes")
return squeeze(z)
def simulate(N, D, S, G, dt):
x0, v0, m = initial_cond(N, D)
pool = Pool()
if COMM_WORLD.Get_rank() == 0:
for s in range(S):
x1, v1 = timestep(x0, v0, G, m, dt, pool)
x0, v0 = x1, v1
else:
pool.wait()
def array(self, N=None, filename=None, component=None, root=0):
"""Dump data to numpy format on root processor for all or one snapshot."""
is_root = comm.Get_rank() == root
size = self.get_total_number_probes() if is_root else len(self)
comp = self.value_size() if component is None else 1
if not N is None:
z = zeros((size, comp))
else:
z = zeros((size, comp, self.number_of_evaluations()))
# Get all values
if len(self) > 0:
if not N is None:
for k in range(comp):
if is_root:
ids = self.get_probe_ids()
z[ids,
k] = self.get_probes_component_and_snapshot(k, N)
else:
z[:, k] = self.get_probes_component_and_snapshot(k, N)
else:
for i, (index, probe) in enumerate(self):
j = index if is_root else i
if not N is None:
z[j, :] = probe.get_probe_at_snapshot(N)
else:
for k in range(self.value_size()):
z[j, k, :] = probe.get_probe_sub(k)
# Collect values on root
recvfrom = comm.gather(len(self), root=root)
if is_root:
for j, k in enumerate(recvfrom):
if comm.Get_rank() != j:
ids = comm.recv(source=j, tag=101)
z0 = comm.recv(source=j, tag=102)
z[ids, :] = z0[:, :]
else:
ids = self.get_probe_ids()
comm.send(ids, dest=root, tag=101)
comm.send(z, dest=root, tag=102)
if is_root:
if filename:
if not N is None:
save(filename + "_snapshot_" + str(N), z)
else:
save(filename + "_all", z)
return squeeze(z)
def exchange_guard_cells( physical_F_left, physical_F_right ):
# MPI exchanges of guard cells
# Send physical cell to left proc
req1 = mpi_comm.isend( physical_F_left,
dest=(mpi_comm.rank-1)%mpi_comm.size )
# Send physical cell to right proc
req2 = mpi_comm.isend( physical_F_right,
dest=(mpi_comm.rank+1)%mpi_comm.size )
# Receive value from right proc
req3 = mpi_comm.irecv( source=(mpi_comm.rank+1)%mpi_comm.size )
# Receive value from left proc
req4 = mpi_comm.irecv( source=(mpi_comm.rank-1)%mpi_comm.size )
# Wait for the processors to finish sending/receiving
req1.wait()
req2.wait()
F_from_right = req3.wait()
F_from_left = req4.wait()
return F_from_left, F_from_right
def all_reduce(tensor: torch.Tensor,
op=ReduceOp.SUM,
comm: MPI.COMM_WORLD = None) -> torch.Tensor:
param_numpy = tensor.numpy()
param_output = np.empty(param_numpy.shape, dtype=param_numpy.dtype)
if comm is None:
comm = _get_comm()
comm.Allreduce(param_numpy, param_output, op=op.value)
tensor = torch.from_numpy(param_output)
return tensor
def _build_cat_distributed(comm, name, path):
# Control flow explanation:
# * `build_err` starts out as `None`
# * Rank 1 to N wait for a broadcast from rank 0 to receive the new value
# for `build_err`
# * Rank 0 splits off from the others and executes the build.
# * If it builds correctly it finishes the collective `build_err`
# broadcast with the initial value `None`: all nodes continue.
# * If it errors, it finishes the collective broadcast with the caught err
#
# All MPI ranks either continue or raise the same err. (prevents stalling)
build_err = None
if not comm.Get_rank():
try:
_build_cat_local(name, path)
except Exception as e:
build_err = e
build_err = comm.bcast(build_err, root=0)
if build_err:
raise build_err
def test_partition(self):
""" Make sure the x_ranges span the entire space without any gaps. """
shapes = ((200,30,10), (33,10,10), (130,5,5), (111,2,2))
for shape in shapes:
space.initialize_space(shape)
x = comm.gather(space.get_space_info()['x_range'])
if comm.Get_rank() == 0:
self.assertEqual(x[0][0], 0)
self.assertEqual(x[-1][-1], space.get_space_info()['shape'][0])
for k in range(len(x)-1):
self.assertEqual(x[k][1], x[k+1][0])
def __init__(self):
"""Obtain configurations and all filelists from the master node
"""
assert world.rank >= 1
# synchronize configuration file and file lists with the master node
self.config = {}
self.mapIn = []
self.reduceIn = []
self.reduceOut = []
self.config = world.bcast(self.config, root=0)
def run(self):
"""Running sequence of the master node
"""
atimer = Timer('Master')
while self.iterCtrl():
# map phase
if self.init:
self.execTask('Init')
self.init = False
else:
self.execTask('Map')
# reduce phase
self.execTask('Reduce')
# terminate all workers
for ii in range(1, len(self.workers)+1):
world.send(True, dest=ii, tag=TERMINATE)
# final output, serial execution only
self.finalize()
def run(self):
"""Running sequence of the master node
"""
atimer = Timer('Master')
while self.iterCtrl():
# map phase
if self.init:
self.execTask('Init')
self.init = False
else:
self.execTask('Map')
# reduce phase
self.execTask('Reduce')
# terminate all workers
for ii in range(1, len(self.workers) + 1):
world.send(True, dest=ii, tag=TERMINATE)
# final output, serial execution only
self.finalize()
def test_recover(self):
""" Make sure we can store and retrieve information from the GPU. """
for case in self.cases:
space.initialize_space(case['shape'])
data = np.random.randn(*case['shape']).astype(case['dtype'])
cpu_data = np.empty_like(data)
comm.Allreduce(data, cpu_data)
g = Grid(cpu_data)
gpu_data = g.get()
if comm.Get_rank() == 0:
self.assertTrue((cpu_data == gpu_data).all())
# Test with-overlap cases as well.
for k in range(1, 3):
g = Grid(cpu_data, x_overlap=k)
gpu_data = g.get()
if comm.Get_rank() == 0:
self.assertTrue((cpu_data == gpu_data).all())
cpu_raw = get_cpu_raw(cpu_data, k)
self.assertTrue((cpu_raw == g._get_raw()).all())
def plot_fields( self, save_figure=False ):
"""
Plot the Ex and By field using matplotlib
If save_figure is True, the plots are saved as PNG files,
in a folder named `diagnostics`
"""
# PLOTTING: NEW LINES RELATED TO MPI
Ex_from_all_procs = mpi_comm.gather( self.Ex[1:-1] )
By_from_all_procs = mpi_comm.gather( self.By[1:-1] )
if mpi_comm.rank == 0:
print( 'Plotting the fields at iteration %d' %self.n )
global_Ex = np.concatenate( Ex_from_all_procs )
global_By = np.concatenate( By_from_all_procs )
plt.clf()
plt.suptitle('Fields at iteration %d' %self.n)
# Plot of Ex
plt.subplot(211)
z = self.dz*np.arange( self.Nz_global )
plt.plot( z, global_Ex, '-' )
plt.ylim(-1.1, 1.1)
plt.xlim(0, self.Lz)
plt.ylabel('$E_x^n$')
plt.xlabel('z')
# Plot of By
plt.subplot(212)
z = self.dz*np.arange( self.Nz_global ) + 0.5*self.dz
plt.plot( z, global_By, '-' )
plt.ylim(-1.1/c, 1.1/c)
plt.xlim(0, self.Lz)
plt.ylabel('$B_y^{n-1/2}$')
plt.xlabel('z')
if save_figure is True:
# Check that the diagnostics folder exists
if os.path.exists('diagnostics') is False:
os.mkdir('diagnostics')
plt.savefig( "diagnostics/iteration%03d.png" %self.n)
def all_reduce(tensor: EagerTensor,
op=ReduceOp.SUM,
comm: MPI.COMM_WORLD = None) -> EagerTensor:
param_numpy = tensor.numpy()
original_shape = param_numpy.shape
param_numpy_flatten = param_numpy.flatten()
param_output = np.empty(param_numpy_flatten.shape, dtype=param_numpy.dtype)
if comm is None:
comm = _get_comm()
comm.Allreduce(param_numpy_flatten, param_output, op=op.value)
param_output = np.reshape(param_output, original_shape)
tensor = tutils.to_tensor(param_output)
return tensor
def test_simple_kernel(self):
""" Implement a simple kernel. """
for case in self.cases:
# Form data to work on.
space.initialize_space(case['shape'])
x_np = comm.allreduce(np.random.randn(*case['shape']).astype(case['dtype']))
x = Grid(x_np, x_overlap=2)
s_np = comm.allreduce(np.random.randn(case['shape'][0],1,1).astype(case['dtype']))
s = Const(s_np)
z = Out(case['dtype'])
# Make a kernel.
code = Template("""
if (_in_local && _in_global) {
z += a * s(_X) * x(0,0,0);
// z += a * x(0,0,0);
}
""").render()
fun = Kernel(code, \
('a', 'number', case['dtype']), \
('x', 'grid', x.dtype), \
('s', 'const', s.dtype), \
('z', 'out', z.dtype), \
shape_filter='all')
# Execute and check the result.
# fun()
while fun.exec_configs:
# for k in range(40):
fun(case['dtype'](2.0), x, s, z)
# fun(case['dtype'](2.0), x, z)
gpu_sum = z.get()
cpu_sum = np.sum(2 * s_np * x_np)
# cpu_sum = np.sum(2 * x_np)
err = abs(gpu_sum - cpu_sum) / abs(cpu_sum)
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-2, (case, err))
else:
self.assertTrue(err < 1e-6, (case, err))
def plot(x, psi, psi_0, nt, v):
###
rcParams["figure.figsize"] = [8 / mpi.Get_size(), 5]
###
pyplot.step(x, psi_0(x), label='initial', where='mid')
pyplot.step(x, psi_0(x - v * nt), label='analytical', where='mid')
pyplot.step(x, psi, label='numerical', where='mid')
pyplot.grid()
pyplot.gca().set_ylim([0, 12])
pyplot.legend()
###
# pyplot.savefig("out.svg")
pyplot.savefig(f"out.{mpi.Get_rank()}.svg")
def __init__(self, config):
"""Read in user created config module and initialize the master node
"""
# set default values and update according to user input (config)
# NOTE: input files should be prepared in the user module
# (e.g. split the BIG file into smaller chunks using 'split')
# each file is fed into a mapper, supposing it can fit into mapper's memory
assert hasattr(config, 'mapfn') and hasattr(config, 'reducefn')
self.config = {'nReduce':1, 'nMap':1, 'maxLoop':1, 'appendReduce':True,\
'scratchFolder':'./', 'readPickle':False, 'writePickle':False,\
'verbosity':6, 'timeout':60, 'delay':0.2, 'jobwait':1,\
'mapfn':config.mapfn, 'reducefn':config.reducefn, 'ctrlfn':None,\
'finalfn':None, 'readfn':None, 'hashfn':hash }
if world.size == 1:
raise AttributeError('Parallel mode only! At least one worker node is required.')
# number of mapping tasks by default equals number of initial files
# it can be overidden by user input
assert isinstance(config.initFiles, list)
self.config['nMap'] = len(config.initFiles)
self.initFiles = config.initFiles
# read in user defined configurations
for key, val in self.config.items():
if hasattr(config, key): self.config[key] = getattr(config, key)
# sync config with all nodes
self.config = world.bcast(self.config, root=0)
# setup workers into a priority queue
self.workers = [ State(ii) for ii in range(1, world.size) ]
heapq.heapify(self.workers)
self.nActive = world.size - 1
# assign map / reduce / finalize file list
tmpList = [ config.__name__+'_'+str(ii).zfill(len(str(self.config['nMap'])))\
+'.map' for ii in range(1, self.config['nMap']+1) ]
self.mapIn = [ os.path.join(self.config['scratchFolder'], file) for file in tmpList ]
tmpList = [ config.__name__+'_'+str(ii).zfill(len(str(self.config['nReduce'])))\
+'.int' for ii in range(1, self.config['nReduce']+1) ]
self.reduceIn = [ os.path.join(self.config['scratchFolder'], file) for file in tmpList ]
self.reduceOut = [ os.path.splitext(file)[0]+'.red' for file in self.reduceIn ]
# Currently only support single output file
self.finalOut = [ config.__name__+'.out' ]
# count number of iterations
self.nLoop = 0; self.init = True
def execTask(self, task):
"""Wrapper function calling mapping/reducing/finalizing phase tasks,
dispatch tasks to workers until all finished and collect feedback.
Faulty workers are removed from active duty work list.
"""
atimer = Timer(task)
print('Entering {0:s} phase...'.format(task))
taskDict = { 'Map':(self.mapIn, MAP_START, MAP_FINISH), \
'Init':(self.mapIn, INIT_START, MAP_FINISH), \
'Reduce':(self.reduceIn, REDUCE_START, REDUCE_FINISH) }
# line up jobs and workers into priority queues
jobs = taskDict[task][0][:]
heapq.heapify(jobs)
running = {}
heapq.heapify(self.workers)
while (jobs or running) and self.nActive > 0:
# dispatch all jobs to all free workers
while jobs and self.workers[0].isFree():
job = heapq.heappop(jobs)
worker = heapq.heappop(self.workers)
world.send(job, dest=worker.id, tag=taskDict[task][1])
worker.setBusy()
heapq.heappush(self.workers, worker)
running[job] = (time(), worker)
if self.config['verbosity'] >= 6:
print('Dispatching file ' + os.path.basename(job) +
' to worker ' + str(worker.id))
# if no more free workers, break
if not self.workers[0].isFree(): break
# wait for finishing workers as well as do cleaning
self.wait(running, taskDict[task][2])
self.clean(running, jobs)
print('{0:s} phase completed'.format(task))
def reduce(self, tag):
"""Use supplied reducefn to operate on
a list of values from a given key, generated by self.map()
"""
atimer = Timer('Worker_Reduce')
filename = world.recv(source=0, tag=tag)
files = glob.glob(filename + '-tmp*')
dataList = []
for file in files:
with open(file, 'rb') as fin:
try:
while True:
dataList.extend(pickle.load(fin))
except EOFError: # read in every instance of pickle dump
pass
data = {}
for key, val in dataList:
if key in data:
data[key].append(val)
else:
data[key] = [val]
results = []
for key, values in data.items():
results.append((key, self.config['reducefn'](key, values)))
results.sort(key=itemgetter(0))
# write out in dictionary format
idx = self.reduceIn.index(filename)
if self.config['appendReduce']:
with open(self.reduceOut[idx], 'a+') as fout:
pickle.dump(dict(results), fout, pickle.HIGHEST_PROTOCOL)
else:
with open(self.reduceOut[idx], 'w+') as fout:
pickle.dump(dict(results), fout, pickle.HIGHEST_PROTOCOL)
world.send(filename, dest=0, tag=REDUCE_FINISH)
def process_dir(indir, outdir):
main_text_files = glob.glob("{0}/main/*.txt".format(indir))
rank = world.Get_rank()
size = world.Get_size()
main_text_files_2 = []
for m in main_text_files:
tilename = find_tilename(m)
out_main = "{0}/main/{1}.fits".format(outdir, tilename)
out_epoch = "{0}/epoch/{1}.fits".format(outdir, tilename)
if not (os.path.exists(out_main) and os.path.exists(out_epoch)):
main_text_files_2.append(m)
main_text_files = main_text_files_2
print "{0} files left to do".format(len(main_text_files))
for i, main_text_file in enumerate(main_text_files):
if i % size != rank:
continue
print rank, main_text_file
tilename = find_tilename(main_text_file)
epoch_text_file = "{0}/epoch/{1}.epoch.txt".format(indir, tilename)
out_main = "{0}/main/{1}.fits".format(outdir, tilename)
out_epoch = "{0}/epoch/{1}.fits".format(outdir, tilename)
if os.path.exists(out_main) and os.path.exists(out_epoch):
continue
try:
process_text(main_text_file,
epoch_text_file,
out_main,
out_epoch,
"r",
blind=False,
quiet=False,
report=report)
except:
print "{} did not work".format(out_main)
if report:
return
def test_synchronize(self):
""" Make sure that we can make the overlap spaces accurate. """
for case in self.cases:
space.initialize_space(case['shape'])
data = np.random.randn(*case['shape']).astype(case['dtype'])
cpu_data = np.empty_like(data)
comm.Allreduce(data, cpu_data)
g = Grid(case['dtype'])
self.assertRaises(TypeError, g.synchronize) # No overlap.
# Test with-overlap cases as well.
for k in range(1, 4):
g = Grid(case['dtype'], x_overlap=k)
# Overwrite entire grid
data = np.random.randn(*case['shape']).astype(case['dtype'])
cpu_data = np.empty_like(data)
comm.Allreduce(data, cpu_data)
cpu_raw_bad = get_cpu_raw(cpu_data, k)
cpu_raw_bad[:k, :, :] += 1 # Mess up padding areas.
cpu_raw_bad[-k:, :, :] += 1
drv.memcpy_htod(g.data.ptr, cpu_raw_bad)
# Prove that the data is not synchronized at this time.
cpu_raw = get_cpu_raw(cpu_data, k)
xx = case['shape'][0]
gd = g._get_raw()
self.assertTrue((gd[:k, :, :] != cpu_raw[:k, :, :]).all())
self.assertTrue((gd[-k:, :, :] != cpu_raw[-k:, :, :]).all())
g.synchronize() # Synchronize the overlapping data.
# Make sure that the overlap data is accurate.
gd = g._get_raw()
self.assertTrue((gd[:k, :, :] == cpu_raw[:k, :, :]).all())
self.assertTrue((gd[-k:, :, :] == cpu_raw[-k:, :, :]).all())
comm.Barrier() # Wait for other mpi nodes to finish.
def main(nt, nx, dt, C, x_min, x_max):
dx = (x_max - x_min) / nx
###
size = mpi.Get_size()
rank = mpi.Get_rank()
# dla nx=5 i size=3: lepiej 2+2+1 niż 1+1+3
import math
nx_max = math.ceil(nx / size)
nx = nx_max if (rank + 1) * nx_max <= nx else nx - rank * nx_max
assert nx > 0
x_min += dx * nx_max * rank
x_max = min(x_max, x_min + dx * nx_max)
#print(rank, '/', size, ':', nx, x_min, x_max)
###
x = np.linspace(x_min - halo * dx,
x_max + halo * dx,
num=nx + 2 * halo,
endpoint=False)
psi = calc(psi_0(x), nt, C)
plot(x[halo:-halo], psi[halo:-halo], psi_0, nt, v=C / dt * dx)
def reduce(self, tag):
"""Use supplied reducefn to operate on
a list of values from a given key, generated by self.map()
"""
atimer = Timer('Worker_Reduce')
filename = world.recv(source=0, tag=tag)
files = glob.glob(filename+'-tmp*')
dataList = []
for file in files:
with open(file, 'rb') as fin:
try:
while True: dataList.extend( pickle.load(fin) )
except EOFError: # read in every instance of pickle dump
pass
data = {}
for key, val in dataList:
if key in data:
data[key].append(val)
else:
data[key] = [val]
results = []
for key, values in data.items():
results.append( ( key, self.config['reducefn'](key, values) ) )
results.sort(key=itemgetter(0))
# write out in dictionary format
idx = self.reduceIn.index(filename)
if self.config['appendReduce']:
with open(self.reduceOut[idx], 'a+') as fout:
pickle.dump(dict(results), fout, pickle.HIGHEST_PROTOCOL)
else:
with open(self.reduceOut[idx], 'w+') as fout:
pickle.dump(dict(results), fout, pickle.HIGHEST_PROTOCOL)
world.send(filename, dest=0, tag=REDUCE_FINISH)
def array(self, N=None, filename=None, component=None, root=0):
"""Dump data to numpy format on root processor for all or one snapshot."""
is_root = comm.Get_rank() == root
size = self.get_total_number_probes() if is_root else len(self)
comp = self.value_size() if component is None else 1
if not N is None:
z = zeros((size, comp))
else:
z = zeros((size, comp, self.number_of_evaluations()))
# Get all values
if len(self) > 0:
if not N is None:
for k in range(comp):
if is_root:
ids = self.get_probe_ids()
z[ids, k] = self.get_probes_component_and_snapshot(k, N)
else:
z[:, k] = self.get_probes_component_and_snapshot(k, N)
else:
for i, (index, probe) in enumerate(self):
j = index if is_root else i
if not N is None:
z[j, :] = probe.get_probe_at_snapshot(N)
else:
for k in range(self.value_size()):
z[j, k, :] = probe.get_probe_sub(k)
# Collect values on root
recvfrom = comm.gather(len(self), root=root)
if is_root:
for j, k in enumerate(recvfrom):
if comm.Get_rank() != j:
ids = comm.recv(source=j, tag=101)
z0 = comm.recv(source=j, tag=102)
z[ids, :] = z0[:, :]
else:
ids = self.get_probe_ids()
comm.send(ids, dest=root, tag=101)
comm.send(z, dest=root, tag=102)
if is_root:
if filename:
if not N is None:
z.dump(filename+"_snapshot_"+str(N)+".probes")
else:
z.dump(filename+"_all.probes")
return squeeze(z)
def __init__(self, N=(256, ), x0=(0.0, ), x1=(1.0, )):
assert len(N) is len(x0) is len(x1)
try:
from mpi4py.MPI import COMM_WORLD, Compute_dims
except ImportError:
print "Error! DistributedDomain requires the mpi4py package."
exit()
mpi_sizes = Compute_dims(COMM_WORLD.size, len(N))
cart = COMM_WORLD.Create_cart(mpi_sizes, periods=[True for n in N])
mpi_coord = cart.Get_coords(COMM_WORLD.rank)
global_shape = [n for n in N]
global_start = [0 for n in N]
local_shape = []
X0, X1 = [], []
dx = [float(l1 - l0) / n for n, l0, l1 in zip(N, x0, x1)]
for i in range(len(N)):
R = N[i] % mpi_sizes[i]
normal_size = N[i] / mpi_sizes[i]
augmnt_size = normal_size + 1
thisdm_size = augmnt_size if mpi_coord[i] < R else normal_size
for j in range(mpi_coord[i]):
global_start[i] += augmnt_size if j < R else normal_size
local_shape.append(thisdm_size)
X0.append(x0[i] + dx[i] * global_start[i])
X1.append(x0[i] + dx[i] * (global_start[i] + thisdm_size))
self.N = local_shape
self.dx = dx
self.x0 = X0
self.x1 = X1
self.cart = cart
self.rank = COMM_WORLD.rank
self.mpi_coord = mpi_coord
self.mpi_sizes = mpi_sizes
self.global_start = global_start
self.global_shape = global_shape
self.is_distributed = True
def test_sum(self):
""" Make sure summing works. """
for case in self.cases:
space.initialize_space(case['shape'])
x = Out(case['dtype'], op='sum')
x_cpu_data = np.random.randn(*case['shape'][1:]).astype(case['dtype'])
if case['dtype'] in (np.complex64, np.complex128):
x_cpu_data = (1 + 1j) * x_cpu_data
x.data.set(x_cpu_data)
res_gold = comm.allreduce(np.sum(x_cpu_data.flatten()))
x.reduce()
err = abs(res_gold - x.get()) / abs(res_gold)
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-3)
else:
self.assertTrue(err < 1e-10)
def test_sum(self):
""" Make sure summing works. """
for case in self.cases:
space.initialize_space(case['shape'])
x = Out(case['dtype'], op='sum')
x_cpu_data = np.random.randn(*case['shape'][1:]).astype(
case['dtype'])
if case['dtype'] in (np.complex64, np.complex128):
x_cpu_data = (1 + 1j) * x_cpu_data
x.data.set(x_cpu_data)
res_gold = comm.allreduce(np.sum(x_cpu_data.flatten()))
x.reduce()
err = abs(res_gold - x.get()) / abs(res_gold)
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-3)
else:
self.assertTrue(err < 1e-10)
def set_periodic_checkpoint(sim, period):
"""
Set up periodic checkpoints of the simulation
The checkpoints are saved in openPMD format, in the directory
`./checkpoints`, with one subdirectory per process.
All the field and particle information of each processor is saved.
NB: Checkpoints are registered among the list of diagnostics
`diags` of the Simulation object `sim`.
Parameters
----------
sim: a Simulation object
The simulation that is to be saved in checkpoints
period: integer
The number of PIC iteration between each checkpoint.
"""
# Only processor 0 creates a directory where checkpoints will be stored
# Make sure that all processors wait until this directory is created
# (Use the global MPI communicator instead of the `BoundaryCommunicator`
# so that this still works in the case `use_all_ranks=False`)
if comm.rank == 0:
if os.path.exists('./checkpoints') is False:
os.mkdir('./checkpoints')
comm.Barrier()
# Choose the name of the directory: one directory per processor
write_dir = 'checkpoints/proc%d/' % comm.rank
# Register a periodic FieldDiagnostic in the diagnostics of the simulation
sim.diags.append(FieldDiagnostic(period, sim.fld, write_dir=write_dir))
# Register a periodic ParticleDiagnostic, which contains all
# the particles which are present in the simulation
particle_dict = {}
for i in range(len(sim.ptcl)):
particle_dict['species %d' % i] = sim.ptcl[i]
sim.diags.append(
ParticleDiagnostic(period, particle_dict, write_dir=write_dir))
maf.bcast = bcast
def compare(msg, mpi_func, maf_func, count=100):
mpi_times = numpy.zeros(count)
maf_times = numpy.zeros(count)
for i in range(count):
start = timeit.default_timer()
mpi_func()
mpi_times[i] = timeit.default_timer() - start
start = timeit.default_timer()
maf_func()
maf_times[i] = timeit.default_timer() - start
maf.log('{:<20} {:.4e} {:.4e} {:>6.2f}'.format(msg, mpi_times.mean(), maf_times.mean(),
100*((maf_times - mpi_times) / mpi_times).mean()))
compare('bcast(s)', lambda: COMM_WORLD.bcast('s'), lambda: maf.bcast('s'))
compare('bcast(1)', lambda: COMM_WORLD.bcast(1), lambda: maf.bcast(1))
compare('bcast(s*100)', lambda: COMM_WORLD.bcast('s'*100), lambda: maf.bcast('s'*100))
compare('bcast(s*1000)', lambda: COMM_WORLD.bcast('s'*1000), lambda: maf.bcast('s'*1000))
compare('bcast(range(1000))',
lambda: COMM_WORLD.bcast(range(1000)),
lambda: maf.bcast(range(1000)))
compare('bcast(range(10000))',
lambda: COMM_WORLD.bcast(range(10000)),
lambda: maf.bcast(range(10000)))
import pickle
import matplotlib.patches
import collections
import sys
from scipy import stats
from mpi4py.MPI import COMM_WORLD as CW
def flatten(x):
if isinstance(x, collections.Iterable):
return [a for i in x for a in flatten(i)]
else:
return [x]
rank = CW.Get_rank()
size = CW.Get_size()
two_col_width = 7.20472 #inches
single_col_width = 3.50394 #inches
page_height = 10.62472
font_size = 10
sys.stdout.flush()
CW.Barrier()
pickle_file = sys.argv[1]
true_birth_con_pickle = sys.argv[2]
plot_gradient = False
read_pickle = bool(sys.argv[3])
baseline_yr = float(sys.argv[4])
def reduce(self):
""" Compute the result. """
self.result = comm.allreduce(ga.sum(self.data).get())
def batch_reduce(*outs):
""" Optimal (compared to self.reduce) when communication cost is latency bound. """
results = comm.allreduce(np.array([ga.sum(out.data).get() for out in outs]))
for k in range(len(outs)):
outs[k].result = results[k]
def __init__(self, array_or_dtype, x_overlap=0):
""" Create a spatial grid on the GPU(s).
Input variables
array_or_dtype -- can either be a numpy array of the same shape as
the global space, or a numpy dtype. If a valid array is passed,
it will be loaded on to the GPU. If a dtype is passed, then
an array of zeros, of that dtype will be loaded onto the GPU.
Optional variables
x_overlap -- the number of adjacent cells in either the negative or
positive x-direction that need to simultaneously be accessed along
with the current cell. Must be a non-negative integer. Default
value is 0.
"""
shape = get_space_info()['shape'] # Get the shape of the space.
xr = get_space_info()['x_range'] # Get the local x_range.
all_x_ranges = get_space_info()['all_x_ranges'] # Get the local x_range.
local_shape = (xr[1]-xr[0], shape[1], shape[2])
self._set_gce_type('grid') # Set the gce type to grid.
# Make sure overlap option is valid.
if type(x_overlap) is not int:
raise TypeError('x_overlap must be an integer.')
elif x_overlap < 0:
raise TypeError('x_overlap must be a non-negative integer.')
if comm.rank == 0:
# Process the array_or_dtype input variable.
if type(array_or_dtype) is np.ndarray: # Input is an array.
array = array_or_dtype
# Make sure the array is of the correct shape.
if array.shape != shape:
raise TypeError('Shape of array does not match shape of space.')
# Make sure the array is of a valid datatype.
self._get_dtype(array.dtype.type)
elif type(array_or_dtype) is type: # Input is a datatype.
self._get_dtype(array_or_dtype) # Validate the dtype.
array = np.zeros(shape, dtype=self.dtype) # Make a zeros array.
else: # Invalid input.
raise TypeError('Input variable must be a numpy array or dtype')
# Prepare array to be scattered.
array = [array[r[0]:r[1],:,:] for r in all_x_ranges]
else:
array = None
array = comm.scatter(array)
self._get_dtype(array.dtype.type)
# # Narrow down the array to local x_range.
# array = array[xr[0]:xr[1],:,:]
# Add padding to array, if needed.
self._xlap = x_overlap
if self._xlap is not 0:
padding = np.empty((self._xlap,) + shape[1:3], dtype=array.dtype)
array = np.concatenate((padding, array, padding), axis=0)
self.to_gpu(array) # Load onto device.
# Determine information needed for synchronization.
if self._xlap is not 0:
# Calculates the pointer to the x offset in a grid.
ptr_dx = lambda x_pos: self.data.ptr + self.data.dtype.itemsize * \
x_pos * shape[1] * shape[2]
# Pointers to different sections of the grid that are relevant
# for synchronization.
self._sync_ptrs = { 'forw_src': ptr_dx(xr[1]-xr[0]), \
'back_dest': ptr_dx(0), \
'back_src': ptr_dx(self._xlap), \
'forw_dest': ptr_dx(xr[1]-xr[0] + self._xlap)}
# Buffers used during synchronization.
self._sync_buffers = [drv.pagelocked_empty( \
(self._xlap, shape[1], shape[2]), \
self.dtype) for k in range(4)]
# Streams used during synchronization.
self._sync_streams = [drv.Stream() for k in range(4)]
# Used to identify neighboring MPI nodes with whom to synchronize.
self._sync_adj = get_space_info()['mpi_adj']
# Offset in bytes to the true start of the grid.
# This is used to "hide" overlap areas from the kernel.
self._xlap_offset = self.data.dtype.itemsize * \
self._xlap * shape[1] * shape[2]
self.synchronize() # Synchronize the grid.
comm.Barrier() # Wait for all grids to synchronize before proceeding.
def __del__(self):
if self.rank == 0:
for p in range(1, self.P):
COMM_WORLD.isend(False, dest=p)
def get_parameters(name):
""" Reads the simulation parameters from the input hdf5 file. """
if comm.rank == 0:
f = h5py.File(name + '.grid', 'r')
files_to_delete = [name + '.grid']
omega = np.complex128(f['omega_r'][0] + 1j * f['omega_i'][0])
shape = tuple([int(s) for s in f['shape'][:]])
# bound_conds = f['bound_conds'][:]
# Function used to read in a 1D complex vector fields.
get_1D_fields = lambda a: [(f[a+'_'+u+'r'][:] + 1j * f[a+'_'+u+'i'][:]).\
astype(np.complex128) for u in 'xyz']
# Read in s and t vectors.
s = get_1D_fields('sp')
t = get_1D_fields('sd')
# Read in max_iters and err_thresh.
max_iters = int(f['max_iters'][0])
# max_iters = 100
err_thresh = float(f['err_thresh'][0])
f.close() # Close file.
# Function used to read in 3D complex vector fields.
def get_3D_fields(a):
field = []
for k in range(3):
key = name + '.' + a + '_' + 'xyz'[k]
field.append((h5py.File(key + 'r')['data'][:] + \
1j * h5py.File(key + 'i')['data'][:]).astype(np.complex128))
files_to_delete.append(key + 'r')
files_to_delete.append(key + 'i')
return field
# # Read in m, e, and j fields.
# for name in 'eJmE':
# print comm.rank, name
# params[name] = get_3D_fields(name)
e = get_3D_fields('e')
j = get_3D_fields('J')
m = get_3D_fields('m')
x = get_3D_fields('E')
# Delete input files.
for filename in files_to_delete:
os.remove(filename)
# Do some simple pre-computation.
for k in range(3):
m[k] = m[k]**-1
e[k] = omega**2 * e[k]
j[k] = -1j * omega * j[k]
params = {'omega': omega, 'shape': shape, \
'max_iters': max_iters, 'err_thresh': err_thresh, \
's': s, 't': t}
# 'e': e, 'm': m, 'j': j, 'x': x}
else:
params = None
params = comm.bcast(params)
if comm.rank == 0:
params['e'] = e
params['m'] = m
params['j'] = j
params['x'] = x
else:
for field_name in 'emjx':
params[field_name] = [None] * 3
return params
