# want the distributed scheduler
out.compute()
# In[ ]:
debug
# The trouble with this approach is that Dask is meant for the execution of large datasets/computations - you probably can't simply run the whole thing
# in one local thread, else you wouldn't have used Dask in the first place. So the code above should only be used on a small part of the data that also exchibits the error.
# Furthermore, the method will not work when you are dealing with futures (such as `f`, above, or after persisting) instead of delayed-based computations.
#
# As alternative, you can ask the scheduler to analyze your calculation and find the specific sub-task responsible for the error, and pull only it and its dependnecies locally for execution.
# In[17]:
c.recreate_error_locally(f)
# In[ ]:
debug
# Finally, there are errors other than exceptions, when we need to look at the state of the scheduler/workers. In the standard "LocalCluster" we started, we
# have direct access to these.
#
# In[19]:
c.cluster.scheduler.nbytes
# Or we could start ipython processed in remote workers/schedulers to enable examining their states from an interactive session.
def main(args):
"""
Initialising launch sequence.
"""
# ------------------------------------------------------
# Print some stuff to show that the code is running:
print("")
os.system(
"printf 'A demonstration of a \033[5mDPrepB/DPrepC\033[m SDP pipeline\n'"
)
print("")
# Set the directory for the moment images:
MOMENTS_DIR = args.outputs + '/MOMENTS'
# Check that the output directories exist, if not then create:
os.makedirs(args.outputs, exist_ok=True)
os.makedirs(MOMENTS_DIR, exist_ok=True)
# Set the polarisation definition of the instrument:
POLDEF = init_inst(args.inst)
# Setup Variables for SIP services
# ------------------------------------------------------
# Define the Queue Producer settings:
if args.queues:
queue_settings = {
'bootstrap.servers': 'scheduler:9092',
'message.max.bytes': 100000000
} #10.60.253.31:9092
# Setup the Confluent Kafka Queue
# ------------------------------------------------------
if args.queues:
from confluent_kafka import Producer
import pickle
# Create an SDP queue:
sip_queue = Producer(queue_settings)
# Define a Data Array Format
# ------------------------------------------------------
def gen_data(channel):
return np.array([
vis1[channel], vis2[channel], channel, None, None, False, False,
args.plots,
float(args.uvcut),
float(args.pixels), POLDEF, args.outputs,
float(args.angres), None, None, None, None, None, None, args.twod,
npixel_advice, cell_advice
])
# Setup the Dask Cluster
# ------------------------------------------------------
starttime = t.time()
dask.config.set(get=dask.distributed.Client.get)
client = Client(
args.daskaddress) # scheduler for Docker container, localhost for P3.
print("Dask Client details:")
print(client)
print("")
# Define channel range for 1 subband, each containing 40 channels:
channel_range = np.array(range(int(args.channels)))
# Load the data into memory:
"""
The input data should be interfaced with Buffer Management.
"""
print("Loading data:")
print("")
vis1 = [
load('%s/%s' % (args.inputs, args.ms1), range(channel, channel + 1),
POLDEF) for channel in range(0, int(args.channels))
]
vis2 = [
load('%s/%s' % (args.inputs, args.ms2), range(channel, channel + 1),
POLDEF) for channel in range(0, int(args.channels))
]
# Prepare Measurement Set
# ------------------------------------------------------
# Combine MSSS snapshots:
vis_advice = append_visibility(vis1[0], vis2[0])
# Apply a uv-distance cut to the data:
vis_advice = uv_cut(vis_advice, float(args.uvcut))
npixel_advice, cell_advice = uv_advice(vis_advice, float(args.uvcut),
float(args.pixels))
# Begin imaging via the Dask cluster
# ------------------------------------------------------
# Submit data for each channel to the client, and return an image:
# Scatter all the data in advance to all the workers:
"""
The data here could be passed via Data Queues.
Queues may not be ideal. Data throughput challenges.
Need to think more about the optimum approach.
"""
print("Scatter data to workers:")
print("")
big_job = [client.scatter(gen_data(channel)) for channel in channel_range]
# Submit jobs to the cluster and create a list of futures:
futures = [
client.submit(dprepb_imaging, big_job[channel], pure=False, retries=3)
for channel in channel_range
]
"""
The dprepb_imaging function could generate QA, logging, and pass this information via Data Queues.
Queues work well for this.
Python logging calls are preferable. Send them to a text file on the node.
Run another service that watches that file. Or just read from standard out.
The Dockerisation will assist with logs.
"""
print("Imaging on workers:")
# Watch progress:
progress(futures)
# Wait until all futures are complete:
wait(futures)
# Check that no futures have errors, if so resubmit:
for future in futures:
if future.status == 'error':
print("ERROR: Future", future, "has 'error' status, as:")
print(client.recreate_error_locally(future))
print("Rerunning...")
print("")
index = futures.index(future)
futures[index].cancel()
futures[index] = client.submit(dprepb_imaging,
big_job[index],
pure=False,
retries=3)
# Wait until all futures are complete:
wait(futures)
# Gather results from the futures:
results = client.gather(futures, errors='raise')
# Run QA on ARL objects and produce to queue:
if args.queues:
print("Adding QA to queue:")
for result in results:
sip_queue.produce('qa', pickle.dumps(qa_image(result), protocol=2))
sip_queue.flush()
# Return the data element of each ARL object, as a Dask future:
futures = [
client.submit(arl_data_future, result, pure=False, retries=3)
for result in results
]
progress(futures)
wait(futures)
# Calculate the Moment images
# ------------------------------------------------------
# Now use 'distributed Dask arrays' in order to parallelise the Moment image calculation:
# Construct a small Dask array for every future:
print("")
print("Calculating Moment images:")
print("")
arrays = [
da.from_delayed(future,
dtype=np.dtype('float64'),
shape=(1, 4, 512, 512)) for future in futures
]
# Stack all small Dask arrays into one:
stack = da.stack(arrays, axis=0)
# Combine chunks to reduce overhead - is initially (40, 1, 4, 512, 512):
stack = stack.rechunk((1, 1, 4, 64, 64))
# Spread the data around on the cluster:
stack = client.persist(stack)
# Data is now coordinated by the single logical Dask array, 'stack'.
# Save the Moment images:
"""
The output moment images should be interfaced with Buffer Management.
Need to know more about the Buffer specification.
Related to initial data distribution also/staging.
"""
print("Saving Moment images to disk:")
print("")
# First generate a template:
image_template = import_image_from_fits('%s/imaging_dirty_WStack-%s.fits' %
(args.outputs, 0))
# Output mean images:
# I:
image_template.data = stack[:, :, 0, :, :].mean(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template,
'%s/Mean-%s.fits' % (MOMENTS_DIR, 'I'))
# Q:
image_template.data = stack[:, :, 1, :, :].mean(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template,
'%s/Mean-%s.fits' % (MOMENTS_DIR, 'Q'))
# U:
image_template.data = stack[:, :, 2, :, :].mean(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template,
'%s/Mean-%s.fits' % (MOMENTS_DIR, 'U'))
# P:
image_template.data = da.sqrt(
(da.square(stack[:, :, 1, :, :]) +
da.square(stack[:, :, 2, :, :]))).mean(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template,
'%s/Mean-%s.fits' % (MOMENTS_DIR, 'P'))
# Output standard deviation images:
# I:
image_template.data = stack[:, :, 0, :, :].std(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template, '%s/Std-%s.fits' % (MOMENTS_DIR, 'I'))
# Q:
image_template.data = stack[:, :, 1, :, :].std(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template, '%s/Std-%s.fits' % (MOMENTS_DIR, 'Q'))
# U:
image_template.data = stack[:, :, 2, :, :].std(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template, '%s/Std-%s.fits' % (MOMENTS_DIR, 'U'))
# P:
image_template.data = da.sqrt(
(da.square(stack[:, :, 1, :, :]) +
da.square(stack[:, :, 2, :, :]))).std(axis=0).compute()
# Run QA on ARL objects and produce to queue:
if args.queues:
sip_queue.produce('qa',
pickle.dumps(qa_image(image_template), protocol=2))
# Export the data to disk:
export_image_to_fits(image_template, '%s/Std-%s.fits' % (MOMENTS_DIR, 'P'))
# Flush queue:
if args.queues:
sip_queue.flush()
# Make a tarball of moment images:
subprocess.call([
'tar', '-cvf',
'%s/moment.tar' % (MOMENTS_DIR),
'%s/' % (MOMENTS_DIR)
])
subprocess.call(['gzip', '-9f', '%s/moment.tar' % (MOMENTS_DIR)])
endtime = t.time()
print(endtime - starttime)