def parallel_map(client, task, args, message, batchsize=1, background=False):
"""
Helper to map a function over a sequence of inputs, in parallel, with progress meter.
:param client: IPython.parallel.Client instance
:param task: Function
:param args: Must be a list of tuples of arguments that the task function will be mapped onto.
If the function takes a single argument, it still must be a 1-tuple.
:param message: String for progress bar
:param batchsize: Jobs are shipped in batches of this size. Higher numbers mean less network traffic,
but longer execution time per job.
:return: IPython.parallel.AsyncMapResult
"""
njobs = len(args)
nproc = len(client)
view = client.load_balanced_view()
message += ' ({} proc)'.format(nproc)
pbar = setup_progressbar(message, njobs, simple_progress=True)
if not background:
pbar.start()
map_result = view.map(task, *list(zip(*args)), chunksize=batchsize)
if background:
return map_result, client
while not map_result.ready():
map_result.wait(1)
pbar.update(map_result.progress * batchsize)
pbar.finish()
return map_result
def _generic_sequential_matrix_task(task, trees, normalise):
jobs = itertools.combinations(trees, 2)
n_jobs = int(math.ceil(len(trees) * (len(trees)-1) / 2))
pbar = setup_progressbar("Getting inter-tree distances (seq)", n_jobs)
pbar.start()
results = []
for i, (t1, t2, normalise) in enumerate((t1, t2, normalise) for (t1, t2) in jobs):
results.append(task(t1, t2, normalise))
pbar.update(i)
pbar.finish()
return scipy.spatial.distance.squareform(results)
def _generic_async_matrix_task(task, trees, normalise, batch_size=100):
jobs = itertools.combinations(trees, 2)
n_jobs = int(math.ceil(len(trees) * (len(trees)-1) / (2*batch_size)))
pbar = setup_progressbar("Getting inter-tree distances (async)", n_jobs, simple_progress=True)
pbar.start()
# Split the work into batches of 'batch_size'. Each batch is executed asynchronously
# -- this should work better for large amounts of quick tasks
job_chunks = task.chunks(((t1, t2, normalise) for (t1, t2) in jobs), batch_size).group()()
while not job_chunks.ready():
pbar.update(job_chunks.completed_count())
time.sleep(2)
pbar.finish()
results = job_chunks.get()
return scipy.spatial.distance.squareform(flatten_list(results))
def sequential_map(task, args, message):
"""
Helper to map a function over a sequence of inputs, sequentially, with progress meter.
:param client: IPython.parallel.Client instance
:param task: Function
:param args: Must be a list of tuples of arguments that the task function will be mapped onto.
If the function takes a single argument, it still must be a 1-tuple.
:param message: String for progress bar
:param batchsize: Jobs are shipped in batches of this size. Higher numbers mean less network traffic,
but longer execution time per job.
:return: IPython.parallel.AsyncMapResult
"""
njobs = len(args)
pbar = setup_progressbar(message, njobs, simple_progress=True)
pbar.start()
map_result = []
for (i, arglist) in enumerate(args):
map_result.append(task(*arglist))
pbar.update(i)
pbar.finish()
return map_result
