def on_running(self):
"""
Log memory consumption as the computation goes on; it only works
when the environment variable OQ_NO_DISTRIBUTE is set, since it
is intended for debugging purposes.
"""
if no_distribute():
logs.LOG.warn('PyMem: %d mb, PgMem: %d mb' % self.mem_peaks)
def map_reduce(task, task_args, agg, acc):
"""
Given a task and an iterable of positional arguments, apply the
task function to the arguments in parallel and return an aggregate
result depending on the initial value of the accumulator
and on the aggregation function. To save memory, the order is
not preserved and there is no list with the intermediated results:
the accumulator is incremented as soon as a task result comes.
NB: if the environment variable OQ_NO_DISTRIBUTE is set the
tasks are run sequentially in the current process and then
map_reduce(task, task_args, agg, acc) is the same as
reduce(agg, itertools.starmap(task, task_args), acc).
Users of map_reduce should be aware of the fact that when
thousands of tasks are spawned and large arguments are passed
or large results are returned they may incur in memory issue:
this is way the calculators limit the queue with the
`concurrent_task` concept.
:param task: a `celery` task callable.
:param task_args: an iterable over positional arguments
:param agg: the aggregation function, (acc, val) -> new acc
:param acc: the initial value of the accumulator
:returns: the final value of the accumulator
"""
if no_distribute():
for the_args in task_args:
result, exctype = safely_call(task.task_func, the_args)
if exctype:
raise RuntimeError(result)
acc = agg(acc, result)
else:
backend = current_app().backend
unpik = 0
job_id = task_args[0][0]
taskname = task.__name__
mon = LightMonitor("unpickling %s" % taskname, job_id, task)
to_send = 0
pickled_args = []
for args in task_args:
piks = pickle_sequence(args)
pickled_args.append(piks)
to_send += sum(len(p) for p in piks)
logs.LOG.info("Sending %dM", to_send / ONE_MB)
taskset = TaskSet(tasks=map(task.subtask, pickled_args))
for task_id, result_dict in taskset.apply_async().iter_native():
check_mem_usage() # log a warning if too much memory is used
result_pik = result_dict["result"]
with mon:
result, exctype = result_pik.unpickle()
if exctype:
raise RuntimeError(result)
unpik += len(result_pik)
acc = agg(acc, result)
del backend._cache[task_id] # work around a celery bug
logs.LOG.info("Unpickled %dM of received data in %s seconds", unpik / ONE_MB, mon.duration)
return acc
def _map_reduce(task_func, task_args, agg, acc):
"""
Given a callable and an iterable of positional arguments, apply the
callable to the arguments in parallel and return an aggregate
result depending on the initial value of the accumulator
and on the aggregation function. To save memory, the order is
not preserved and there is no list with the intermediated results:
the accumulator is incremented as soon as a task result comes.
:param task_func: a `celery` task callable.
:param task_args: an iterable over positional arguments
:param agg: the aggregation function, (acc, val) -> new acc
:param acc: the initial value of the accumulator
:returns: the final value of the accumulator
NB: if the environment variable OQ_NO_DISTRIBUTE is set the
tasks are run sequentially in the current process and then
map_reduce(task_func, task_args, agg, acc) is the same as
reduce(agg, itertools.starmap(task_func, task_args), acc).
Users of map_reduce should be aware of the fact that when
thousands of tasks are spawned and large arguments are passed
or large results are returned they may incur in memory issue:
this is way the calculators limit the queue with the
`concurrent_task` concept.
"""
if no_distribute():
for the_args in task_args:
acc = agg(acc, task_func(*the_args))
else:
taskset = TaskSet(tasks=map(task_func.subtask, task_args))
for result in taskset.apply_async():
if isinstance(result, Exception):
# TODO: kill all the other tasks
raise result
acc = agg(acc, result)
return acc
def _map_reduce(task_func, task_args, agg, acc):
"""
Given a callable and an iterable of positional arguments, apply the
callable to the arguments in parallel and return an aggregate
result depending on the initial value of the accumulator
and on the aggregation function. To save memory, the order is
not preserved and there is no list with the intermediated results:
the accumulator is incremented as soon as a task result comes.
:param task_func: a `celery` task callable.
:param task_args: an iterable over positional arguments
:param agg: the aggregation function, (acc, val) -> new acc
:param acc: the initial value of the accumulator
:returns: the final value of the accumulator
NB: if the environment variable OQ_NO_DISTRIBUTE is set the
tasks are run sequentially in the current process and then
map_reduce(task_func, task_args, agg, acc) is the same as
reduce(agg, itertools.starmap(task_func, task_args), acc).
Users of map_reduce should be aware of the fact that when
thousands of tasks are spawned and large arguments are passed
or large results are returned they may incur in memory issue:
this is way the calculators limit the queue with the
`concurrent_task` concept.
"""
if no_distribute():
for the_args in task_args:
acc = agg(acc, task_func(*the_args))
else:
taskset = TaskSet(tasks=map(task_func.subtask, task_args))
for result in taskset.apply_async():
if isinstance(result, Exception):
# TODO: kill all the other tasks
raise result
acc = agg(acc, result)
return acc
