def test_create_simulate_vis_list(self):
vis_list = simulate_list_serial_workflow(
frequency=self.frequency, channel_bandwidth=self.channel_bandwidth)
assert len(vis_list) == len(self.frequency)
vt = vis_list[0]
assert isinstance(vt, BlockVisibility)
assert vt.nvis > 0
log.info(
"Starting imaging-pipeline with %d MPI processes nfreqwin %d ntimes %d" %
(size, nfreqwin, ntimes))
print("Starting imaging-pipeline with %d MPI processes nfreqwin %d ntimes %d" %
(size, nfreqwin, ntimes),
flush=True)
log.debug('%d: frequency len %d frequency list:' % (rank, len(frequency)))
#print(frequency,flush=True)
if rank == 0:
bvis_list = simulate_list_serial_workflow(
'LOWBD2',
frequency=frequency,
channel_bandwidth=channel_bandwidth,
times=times,
phasecentre=phasecentre,
order='frequency',
rmax=rmax,
format='blockvis')
else:
bvis_list = list()
vis_list = [convert_blockvisibility_to_visibility(bv) for bv in bvis_list]
log.debug('%d: %d elements in vis_list' % (rank, len(vis_list)))
#log.handlers[0].flush()
#print(vis_list
# In[4]:
if rank == 0:
def trial_case(results,
seed=180555,
context='wstack',
nworkers=8,
threads_per_worker=1,
memory=8,
processes=True,
order='frequency',
nfreqwin=7,
ntimes=3,
rmax=750.0,
facets=1,
wprojection_planes=1,
use_dask=True,
use_serial_imaging=True,
flux_limit=0.3,
nmajor=5,
dft_threshold=1.0,
use_serial_clean=True,
write_fits=False):
""" Single trial for performance-timings
Simulates visibilities from GLEAM including phase errors
Makes dirty image and PSF
Runs ICAL pipeline
The results are in a dictionary:
'context': input - a string describing concisely the purpose of the test
'time overall', overall execution time (s)
'time predict', time to execute GLEAM prediction graph
'time invert', time to make dirty image
'time invert graph', time to make dirty image graph
'time ICAL graph', time to create ICAL graph
'time ICAL', time to execute ICAL graph
'context', type of imaging e.g. 'wstack'
'nworkers', number of workers to create
'threads_per_worker',
'nnodes', Number of nodes,
'processes', 'order', Ordering of data_models
'nfreqwin', Number of frequency windows in simulation
'ntimes', Number of hour angles in simulation
'rmax', Maximum radius of stations used in simulation (m)
'facets', Number of facets in deconvolution and imaging
'wprojection_planes', Number of wprojection planes
'vis_slices', Number of visibility slices (per Visibbility)
'npixel', Number of pixels in image
'cellsize', Cellsize in radians
'seed', Random number seed
'dirty_max', Maximum in dirty image
'dirty_min', Minimum in dirty image
'restored_max',
'restored_min',
'deconvolved_max',
'deconvolved_min',
'residual_max',
'residual_min',
'git_info', GIT hash (not definitive since local mods are possible)
:param results: Initial state
:param seed: Random number seed (used in gain simulations)
:param context: imaging context
:param context: Type of context: '2d'|'timeslice'|'wstack'
:param nworkers: Number of dask workers to use
:param threads_per_worker: Number of threads per worker
:param processes: Use processes instead of threads 'processes'|'threads'
:param order: See simulate_list_list_arlexecute_workflow_workflowkflow
:param nfreqwin: See simulate_list_list_arlexecute_workflow_workflowkflow
:param ntimes: See simulate_list_list_arlexecute_workflow_workflowkflow
:param rmax: See simulate_list_list_arlexecute_workflow_workflowkflow
:param facets: Number of facets to use
:param wprojection_planes: Number of wprojection planes to use
:param use_dask: Use dask or immediate evaluation
:return: results dictionary
"""
# Initialise MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
size = comm.Get_size()
def init_logging():
logging.basicConfig(
filename='pipelines_mpi_timings.log',
filemode='w',
format='%(asctime)s,%(msecs)d %(name)s %(levelname)s %(message)s',
datefmt='%H:%M:%S',
level=logging.INFO)
init_logging()
log = logging.getLogger()
# Initialise logging on the workers. This appears to only work using the process scheduler.
def lprint(*args):
log.info(*args)
print(*args)
lprint("Starting pipelines_mpi_timings")
numpy.random.seed(seed)
results['seed'] = seed
start_all = time.time()
results['context'] = context
results['hostname'] = socket.gethostname()
results['git_hash'] = git_hash()
results['epoch'] = time.strftime("%Y-%m-%d %H:%M:%S")
lprint("Context is %s" % context)
results['nworkers'] = nworkers
results['threads_per_worker'] = threads_per_worker
results['processes'] = processes
results['memory'] = memory
results['order'] = order
results['nfreqwin'] = nfreqwin
results['ntimes'] = ntimes
results['rmax'] = rmax
results['facets'] = facets
results['wprojection_planes'] = wprojection_planes
results['dft threshold'] = dft_threshold
results['use_dask'] = use_dask
lprint("At start, configuration is:")
lprint(sort_dict(results))
# Parameters determining scale of simulation.
frequency = numpy.linspace(1.0e8, 1.2e8, nfreqwin)
centre = nfreqwin // 2
if nfreqwin > 1:
channel_bandwidth = numpy.array(nfreqwin *
[frequency[1] - frequency[0]])
else:
channel_bandwidth = numpy.array([1e6])
times = numpy.linspace(-numpy.pi / 4.0, numpy.pi / 4.0, ntimes)
phasecentre = SkyCoord(ra=+0.0 * u.deg,
dec=-40.0 * u.deg,
frame='icrs',
equinox='J2000')
lprint("****** Visibility creation ******")
# Create the empty BlockVisibility's and persist these on the cluster
if rank == 0:
tmp_bvis_list = simulate_list_serial_workflow(
'LOWBD2',
frequency=frequency,
channel_bandwidth=channel_bandwidth,
times=times,
phasecentre=phasecentre,
order=order,
format='blockvis',
rmax=rmax)
vis_list = [
convert_blockvisibility_to_visibility(bv) for bv in tmp_bvis_list
]
else:
vis_list = list()
#if rank==0:
# import matplotlib.pyplot as plt
# plt.clf()
# plt.hist(vis_list[0].w, bins=100)
# plt.title('Histogram of w samples: rms=%.1f (wavelengths)' % numpy.std(vis_list[0].w))
# plt.xlabel('W (wavelengths)')
# plt.show()
# plt.clf()
# plt.hist(vis_list[0].uvdist, bins=100)
# plt.title('Histogram of uvdistance samples')
# plt.xlabel('UV Distance (wavelengths)')
# plt.show()
sub_vis_list = numpy.array_split(vis_list, size)
sub_vis_list = comm.scatter(sub_vis_list, root=0)
#NOTE: sub_vis_list is scattered (future_vis_list in arlexecute)
# Find the best imaging parameters but don't bring the vis_list back here
print("****** Finding wide field parameters ******")
sub_advice = [
advise_wide_field(v,
guard_band_image=6.0,
delA=0.1,
facets=facets,
wprojection_planes=wprojection_planes,
oversampling_synthesised_beam=4.0)
for v in sub_vis_list
]
# MONTSE: This does not make sense cause they all compute advice but only
# the last one is used. Confirm that this is what the dask version does!!!
advice_list = comm.gather(sub_advice, root=0)
if rank == 0:
advice = numpy.concatenate(advice_list)[-1]
else:
advice = None
advice = comm.bcast(advice, root=0)
# Deconvolution via sub-images requires 2^n
npixel = advice['npixels2']
results['npixel'] = npixel
cellsize = advice['cellsize']
results['cellsize'] = cellsize
lprint("Image will have %d by %d pixels, cellsize = %.6f rad" %
(npixel, npixel, cellsize))
# NOTE: frequency and channel_bandwidth are replicated
sub_frequency = numpy.array_split(frequency, size)
sub_channel_bandwidth = numpy.array_split(channel_bandwidth, size)
# Create an empty model image
tmp_model_list = [
create_image(npixel=npixel,
cellsize=cellsize,
frequency=[sub_frequency[rank][f]],
channel_bandwidth=[sub_channel_bandwidth[rank][f]],
phasecentre=phasecentre,
polarisation_frame=PolarisationFrame("stokesI"))
for f, freq in enumerate(sub_frequency[rank])
]
model_list = comm.gather(tmp_model_list, root=0)
if rank == 0:
model_list = numpy.concatenate(model_list).tolist()
else:
model_list = list()
print(type(model_list))
assert isinstance(model_list, list), model_list
# NOTE: tmp_model_list is the scattered model_list
# NOTE: template_model and gcfcf are replicated
lprint("****** Setting up imaging parameters ******")
# Now set up the imaging parameters
template_model = create_image(
npixel=npixel,
cellsize=cellsize,
frequency=[frequency[centre]],
phasecentre=phasecentre,
channel_bandwidth=[channel_bandwidth[centre]],
polarisation_frame=PolarisationFrame("stokesI"))
gcfcf = [create_pswf_convolutionfunction(template_model)]
if context == 'timeslice':
vis_slices = ntimes
lprint("Using timeslice with %d slices" % vis_slices)
elif context == '2d':
vis_slices = 1
elif context == "wprojection":
wstep = advice['wstep']
nw = advice['wprojection_planes']
vis_slices = 1
support = advice['nwpixels']
results['wprojection_planes'] = nw
lprint("****** Starting W projection kernel creation ******")
lprint("Using wprojection with %d planes with wstep %.1f wavelengths" %
(nw, wstep))
lprint("Support of wprojection = %d pixels" % support)
gcfcf = [
create_awterm_convolutionfunction(template_model,
nw=nw,
wstep=wstep,
oversampling=4,
support=support,
use_aaf=True)
]
lprint("Size of W projection gcf, cf = %.2E bytes" % get_size(gcfcf))
else:
context = 'wstack'
vis_slices = advice['vis_slices']
lprint("Using wstack with %d slices" % vis_slices)
results['vis_slices'] = vis_slices
# Make a skymodel from gleam, with bright sources as components and weak sources in an image
lprint("****** Starting GLEAM skymodel creation ******")
future_skymodel_list = [
create_low_test_skymodel_from_gleam(
npixel=npixel,
cellsize=cellsize,
frequency=[sub_frequency[rank][f]],
phasecentre=phasecentre,
polarisation_frame=PolarisationFrame("stokesI"),
flux_limit=flux_limit,
flux_threshold=dft_threshold,
flux_max=5.0) for f, freq in enumerate(sub_frequency[rank])
]
# We use predict_skymodel so that we can use skycomponents as well as images
lprint("****** Starting GLEAM skymodel prediction ******")
#NOTE: future_skymodel_list is a subset of skymodel_list which has not been
#gathered
sub_predicted_vis_list = [
predict_skymodel_list_serial_workflow(sub_vis_list[f],
[future_skymodel_list[f]],
context=context,
vis_slices=vis_slices,
facets=facets,
gcfcf=gcfcf)[0]
for f, freq in enumerate(sub_frequency[rank])
]
predicted_vis_list = comm.gather(sub_predicted_vis_list, root=0)
if rank == 0:
predicted_vis_list = numpy.concatenate(predicted_vis_list)
else:
predicted_vis_list = list()
#NOTE: sub_predicted_vis_list is predicted_vis_list scattered
# Corrupt the visibility for the GLEAM model
lprint("****** Visibility corruption ******")
tmp_corrupted_vis_list = corrupt_list_serial_workflow(
sub_predicted_vis_list, phase_error=1.0, seed=seed)
lprint("****** Weighting and tapering ******")
tmp_corrupted_vis_list = weight_list_serial_workflow(
tmp_corrupted_vis_list, tmp_model_list)
tmp_corrupted_vis_list = taper_list_serial_workflow(
tmp_corrupted_vis_list, 0.003 * 750.0 / rmax)
corrupted_vis_list = comm.gather(tmp_corrupted_vis_list, root=0)
if rank == 0:
corrupted_vis_list = numpy.concatenate(corrupted_vis_list)
else:
corrupted_vis_list = list()
#future_corrupted_vis_list = arlexecute.scatter(corrupted_vis_list)
# At this point the only futures are of scatter'ed data so no repeated calculations should be
# incurred.
lprint("****** Starting dirty image calculation ******")
start = time.time()
dirty_list = invert_list_mpi_workflow(corrupted_vis_list,
model_list,
vis_slices=vis_slices,
context=context,
facets=facets,
use_serial_invert=use_serial_imaging,
gcfcf=gcfcf)
end = time.time()
results['size invert graph'] = get_size(dirty_list)
lprint('Size of dirty graph is %.3E bytes' %
(results['size invert graph']))
results['time invert graph'] = 0.0
results['time invert'] = end - start
lprint("Dirty image invert took %.3f seconds" % (end - start))
if rank == 0:
dirty, sumwt = dirty_list[centre]
lprint("Maximum in dirty image is %f, sumwt is %s" %
(numpy.max(numpy.abs(dirty.data)), str(sumwt)))
qa = qa_image(dirty)
results['dirty_max'] = qa.data['max']
results['dirty_min'] = qa.data['min']
if write_fits and rank == 0:
export_image_to_fits(
dirty, "pipelines_arlexecute_timings-%s-dirty.fits" % context)
lprint("****** Starting prediction ******")
start = time.time()
result = predict_list_mpi_workflow(corrupted_vis_list,
model_list,
vis_slices=vis_slices,
context=context,
facets=facets,
use_serial_predict=use_serial_imaging,
gcfcf=gcfcf)
# arlexecute.client.cancel(tmp_vis_list)
end = time.time()
results['time predict'] = end - start
lprint("Predict took %.3f seconds" % (end - start))
# Create the ICAL pipeline to run major cycles, starting selfcal at cycle 1. A global solution across all
# frequencies (i.e. Visibilities) is performed.
print("Using subimage clean")
deconvolve_facets = 8
deconvolve_overlap = 16
deconvolve_taper = 'tukey'
lprint("****** Starting ICAL graph creation ******")
controls = create_calibration_controls()
controls['T']['first_selfcal'] = 1
controls['T']['timescale'] = 'auto'
start = time.time()
ical_list = ical_list_mpi_workflow(corrupted_vis_list,
model_imagelist=model_list,
context=context,
vis_slices=vis_slices,
scales=[0, 3, 10],
algorithm='mmclean',
nmoment=3,
niter=1000,
fractional_threshold=0.1,
threshold=0.01,
nmajor=nmajor,
gain=0.25,
psf_support=64,
deconvolve_facets=deconvolve_facets,
deconvolve_overlap=deconvolve_overlap,
deconvolve_taper=deconvolve_taper,
timeslice='auto',
global_solution=True,
do_selfcal=True,
calibration_context='T',
controls=controls,
use_serial_predict=use_serial_imaging,
use_serial_invert=use_serial_imaging,
use_serial_clean=use_serial_clean,
gcfcf=gcfcf)
end = time.time()
results['size ICAL graph'] = get_size(ical_list)
lprint('Size of ICAL graph is %.3E bytes' % results['size ICAL graph'])
results['time ICAL graph'] = 0.0
# Execute the graph
lprint("****** Executing ICAL graph ******")
deconvolved, residual, restored, gaintables = ical_list
results['time ICAL'] = end - start
lprint("ICAL graph execution took %.3f seconds" % (end - start))
if rank == 0:
qa = qa_image(deconvolved[centre])
results['deconvolved_max'] = qa.data['max']
results['deconvolved_min'] = qa.data['min']
deconvolved_cube = image_gather_channels(deconvolved)
if write_fits:
export_image_to_fits(
deconvolved_cube,
"pipelines_arlexecute_timings-%s-ical_deconvolved.fits" %
context)
qa = qa_image(residual[centre][0])
results['residual_max'] = qa.data['max']
results['residual_min'] = qa.data['min']
residual_cube = remove_sumwt(residual)
residual_cube = image_gather_channels(residual_cube)
if write_fits:
export_image_to_fits(
residual_cube,
"pipelines_arlexecute_timings-%s-ical_residual.fits" % context)
qa = qa_image(restored[centre])
results['restored_max'] = qa.data['max']
results['restored_min'] = qa.data['min']
restored_cube = image_gather_channels(restored)
if write_fits:
export_image_to_fits(
restored_cube,
"pipelines_arlexecute_timings-%s-ical_restored.fits" % context)
#
end_all = time.time()
results['time overall'] = end_all - start_all
lprint("At end, results are:")
results = sort_dict(results)
lprint(results)
return results
