def __init__(self, config=None, executors=None, lazy_fail=True,
rundir=None, fail_retries=2):
""" Initialize the DataFlowKernel
Please note that keyword args passed to the DFK here will always override
options passed in via the config.
KWargs:
config (Dict) : A single data object encapsulating all config attributes
executors (list of Executor objs): Optional, kept for (somewhat) backward compatibility with 0.2.0
lazy_fail(Bool) : Default=True, determine failure behavior
rundir (str) : Path to run directory. Defaults to ./runinfo/runNNN
fail_retries(int): Default=2, Set the number of retry attempts in case of failure
Returns:
DataFlowKernel object
"""
# Create run dirs for this run
self.rundir = make_rundir(config=config, path=rundir)
# Update config with defaults
self._config = update_config(config, self.rundir)
# Start the anonymized usage tracker and send init msg
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
if self._config :
self._executors_managed = True
# Create the executors
epf = EPF()
self.executors = epf.make(self.rundir, self._config)
# set global vars from config
self.lazy_fail = self._config["globals"].get("lazyFail", lazy_fail)
self.fail_retries = self._config["globals"].get("fail_retries", fail_retries)
self.flowcontrol = FlowControl(self, self._config)
else:
self._executors_managed = False
self.fail_retries = fail_retries
self.lazy_fail = lazy_fail
self.executors = {i:x for i,x in enumerate(executors)}
print("Executors : ", self.executors)
self.flowcontrol = FlowNoControl(self, None)
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
logger.debug("Using executors: {0}".format(self.executors))
atexit.register(self.cleanup)
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html'
)
self._config = config
self.run_dir = make_rundir(config.run_dir)
parsl.set_file_logger("{}/parsl.log".format(self.run_dir),
level=logging.DEBUG)
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
logger.info("Parsl version: {}".format(get_version()))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# ES logging
self.tasks_completed_count = 0
self.tasks_failed_count = 0
self.monitoring_config = config.monitoring_config
if self.monitoring_config is not None and self.monitoring_config.database_type == 'local_database'\
and self.monitoring_config.eng_link is None:
# uses the rundir as the default location.
logger.info(
'Local monitoring database can be found inside the run_dir at: {}'
.format(self.run_dir))
self.monitoring_config.eng_link = "sqlite:///{}".format(
os.path.join(os.path.abspath(self.run_dir), 'monitoring.db'))
if self.monitoring_config is None:
self.db_logger = get_db_logger()
else:
self.db_logger = get_db_logger(
monitoring_config=self.monitoring_config)
self.workflow_name = None
if self.monitoring_config is not None and self.monitoring_config.workflow_name is not None:
self.workflow_name = self.monitoring_config.workflow_name
else:
for frame in inspect.stack():
fname = os.path.basename(str(frame.filename))
parsl_file_names = ['dflow.py']
# Find first file name not considered a parsl file
if fname not in parsl_file_names:
self.workflow_name = fname
break
self.workflow_version = None
if self.monitoring_config is not None and self.monitoring_config.version is not None:
self.workflow_version = self.monitoring_config.version
self.time_began = time.time()
self.time_completed = None
self.run_id = str(uuid4())
self.dashboard = self.monitoring_config.dashboard_link if self.monitoring_config is not None else None
# TODO: make configurable
logger.info("Run id is: " + self.run_id)
if self.dashboard is not None:
logger.info("Dashboard is found at " + self.dashboard)
# start tornado logging server
if self.monitoring_config is not None and self.monitoring_config.database_type == 'local_database':
self.logging_server = multiprocessing.Process(
target=logging_server.run,
kwargs={'monitoring_config': self.monitoring_config})
self.logging_server.start()
self.web_app = multiprocessing.Process(
target=index.run,
kwargs={'monitoring_config': self.monitoring_config})
self.web_app.start()
else:
self.logging_server = None
self.web_app = None
workflow_info = {
'python_version': sys.version_info,
'parsl_version': get_version(),
"time_began": str(self.time_began),
'time_completed': str(None),
'run_id': self.run_id,
'workflow_name': self.workflow_name,
'workflow_version': self.workflow_version,
'rundir': self.run_dir,
'tasks_completed_count': self.tasks_completed_count,
'tasks_failed_count': self.tasks_failed_count,
'user': getuser(),
'host': gethostname(),
}
self.db_logger.info("DFK start", extra=workflow_info)
# ES logging end
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self,
memoize=config.app_cache,
checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
data_manager = DataManager(
max_threads=config.data_management_max_threads,
executors=config.executors)
self.executors = {
e.label: e
for e in config.executors + [data_manager]
}
for executor in self.executors.values():
executor.run_dir = self.run_dir
if hasattr(executor, 'provider'):
if hasattr(executor.provider, 'script_dir'):
executor.provider.script_dir = os.path.join(
self.run_dir, 'submit_scripts')
if executor.provider.channel.script_dir is None:
executor.provider.channel.script_dir = os.path.join(
self.run_dir, 'submit_scripts')
if not executor.provider.channel.isdir(self.run_dir):
parent, child = pathlib.Path(
self.run_dir).parts[-2:]
remote_run_dir = os.path.join(parent, child)
executor.provider.channel.script_dir = os.path.join(
remote_run_dir, 'remote_submit_scripts')
executor.provider.script_dir = os.path.join(
self.run_dir, 'local_submit_scripts')
executor.provider.channel.makedirs(
executor.provider.channel.script_dir, exist_ok=True)
os.makedirs(executor.provider.script_dir, exist_ok=True)
executor.start()
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception:
logger.error(
"invalid checkpoint_period provided:{0} expected HH:MM:SS".
format(config.checkpoint_period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.tasks = {}
self.submitter_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
class DataFlowKernel(object):
"""The DataFlowKernel adds dependency awareness to an existing executor.
It is responsible for managing futures, such that when dependencies are resolved,
pending tasks move to the runnable state.
Here is a simplified diagram of what happens internally::
User | DFK | Executor
----------------------------------------------------------
| |
Task-------+> +Submit |
App_Fu<------+--| |
| Dependencies met |
| task-------+--> +Submit
| Ex_Fu<------+----|
"""
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html'
)
self._config = config
self.run_dir = make_rundir(config.run_dir)
parsl.set_file_logger("{}/parsl.log".format(self.run_dir),
level=logging.DEBUG)
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
logger.info("Parsl version: {}".format(get_version()))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# ES logging
self.tasks_completed_count = 0
self.tasks_failed_count = 0
self.monitoring_config = config.monitoring_config
if self.monitoring_config is not None and self.monitoring_config.database_type == 'local_database'\
and self.monitoring_config.eng_link is None:
# uses the rundir as the default location.
logger.info(
'Local monitoring database can be found inside the run_dir at: {}'
.format(self.run_dir))
self.monitoring_config.eng_link = "sqlite:///{}".format(
os.path.join(os.path.abspath(self.run_dir), 'monitoring.db'))
if self.monitoring_config is None:
self.db_logger = get_db_logger()
else:
self.db_logger = get_db_logger(
monitoring_config=self.monitoring_config)
self.workflow_name = None
if self.monitoring_config is not None and self.monitoring_config.workflow_name is not None:
self.workflow_name = self.monitoring_config.workflow_name
else:
for frame in inspect.stack():
fname = os.path.basename(str(frame.filename))
parsl_file_names = ['dflow.py']
# Find first file name not considered a parsl file
if fname not in parsl_file_names:
self.workflow_name = fname
break
self.workflow_version = None
if self.monitoring_config is not None and self.monitoring_config.version is not None:
self.workflow_version = self.monitoring_config.version
self.time_began = time.time()
self.time_completed = None
self.run_id = str(uuid4())
self.dashboard = self.monitoring_config.dashboard_link if self.monitoring_config is not None else None
# TODO: make configurable
logger.info("Run id is: " + self.run_id)
if self.dashboard is not None:
logger.info("Dashboard is found at " + self.dashboard)
# start tornado logging server
if self.monitoring_config is not None and self.monitoring_config.database_type == 'local_database':
self.logging_server = multiprocessing.Process(
target=logging_server.run,
kwargs={'monitoring_config': self.monitoring_config})
self.logging_server.start()
self.web_app = multiprocessing.Process(
target=index.run,
kwargs={'monitoring_config': self.monitoring_config})
self.web_app.start()
else:
self.logging_server = None
self.web_app = None
workflow_info = {
'python_version': sys.version_info,
'parsl_version': get_version(),
"time_began": str(self.time_began),
'time_completed': str(None),
'run_id': self.run_id,
'workflow_name': self.workflow_name,
'workflow_version': self.workflow_version,
'rundir': self.run_dir,
'tasks_completed_count': self.tasks_completed_count,
'tasks_failed_count': self.tasks_failed_count,
'user': getuser(),
'host': gethostname(),
}
self.db_logger.info("DFK start", extra=workflow_info)
# ES logging end
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self,
memoize=config.app_cache,
checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
data_manager = DataManager(
max_threads=config.data_management_max_threads,
executors=config.executors)
self.executors = {
e.label: e
for e in config.executors + [data_manager]
}
for executor in self.executors.values():
executor.run_dir = self.run_dir
if hasattr(executor, 'provider'):
if hasattr(executor.provider, 'script_dir'):
executor.provider.script_dir = os.path.join(
self.run_dir, 'submit_scripts')
if executor.provider.channel.script_dir is None:
executor.provider.channel.script_dir = os.path.join(
self.run_dir, 'submit_scripts')
if not executor.provider.channel.isdir(self.run_dir):
parent, child = pathlib.Path(
self.run_dir).parts[-2:]
remote_run_dir = os.path.join(parent, child)
executor.provider.channel.script_dir = os.path.join(
remote_run_dir, 'remote_submit_scripts')
executor.provider.script_dir = os.path.join(
self.run_dir, 'local_submit_scripts')
executor.provider.channel.makedirs(
executor.provider.channel.script_dir, exist_ok=True)
os.makedirs(executor.provider.script_dir, exist_ok=True)
executor.start()
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception:
logger.error(
"invalid checkpoint_period provided:{0} expected HH:MM:SS".
format(config.checkpoint_period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.tasks = {}
self.submitter_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
def _create_task_log_info(self, task_id, fail_mode=None):
"""
Create the dictionary that will be included in the log.
"""
task_log_info = {
"task_" + k: v
for k, v in self.tasks[task_id].items()
}
task_log_info['run_id'] = self.run_id
task_log_info['task_status_name'] = self.tasks[task_id]['status'].name
task_log_info['tasks_failed_count'] = self.tasks_failed_count
task_log_info['tasks_completed_count'] = self.tasks_completed_count
task_log_info['time_began'] = str(self.time_began)
task_log_info['task_inputs'] = str(self.tasks[task_id]['kwargs'].get(
'inputs', None))
task_log_info['task_outputs'] = str(self.tasks[task_id]['kwargs'].get(
'outputs', None))
task_log_info['task_stdin'] = self.tasks[task_id]['kwargs'].get(
'stdin', None)
task_log_info['task_stdout'] = self.tasks[task_id]['kwargs'].get(
'stdout', None)
if fail_mode is not None:
task_log_info['task_fail_mode'] = fail_mode
return task_log_info
def _count_deps(self, depends):
"""Internal.
Count the number of unresolved futures in the list depends.
"""
count = 0
for dep in depends:
if isinstance(dep, Future):
if not dep.done():
count += 1
return count
@property
def config(self):
"""Returns the fully initialized config that the DFK is actively using.
DO *NOT* update.
Returns:
- config (dict)
"""
return self._config
def handle_exec_update(self, task_id, future):
"""This function is called only as a callback from an execution
attempt reaching a final state (either successfully or failing).
It will launch retries if necessary, and update the task
structure.
Args:
task_id (string) : Task id which is a uuid string
future (Future) : The future object corresponding to the task which
makes this callback
KWargs:
memo_cbk(Bool) : Indicates that the call is coming from a memo update,
that does not require additional memo updates.
"""
try:
res = future.result()
if isinstance(res, RemoteExceptionWrapper):
res.reraise()
except Exception:
logger.exception("Task {} failed".format(task_id))
# We keep the history separately, since the future itself could be
# tossed.
self.tasks[task_id]['fail_history'].append(future._exception)
self.tasks[task_id]['fail_count'] += 1
if not self._config.lazy_errors:
logger.debug("Eager fail, skipping retry logic")
self.tasks[task_id]['status'] = States.failed
if self.monitoring_config is not None:
task_log_info = self._create_task_log_info(
task_id, 'eager')
self.db_logger.info("Task Fail", extra=task_log_info)
return
if self.tasks[task_id]['fail_count'] <= self._config.retries:
self.tasks[task_id]['status'] = States.pending
logger.debug("Task {} marked for retry".format(task_id))
if self.monitoring_config is not None:
task_log_info = self._create_task_log_info(task_id, 'lazy')
self.db_logger.info("Task Retry", extra=task_log_info)
else:
logger.info("Task {} failed after {} retry attempts".format(
task_id, self._config.retries))
self.tasks[task_id]['status'] = States.failed
self.tasks_failed_count += 1
self.tasks[task_id]['time_returned'] = time.time()
if self.monitoring_config is not None:
task_log_info = self._create_task_log_info(task_id, 'lazy')
self.db_logger.info("Task Retry Failed",
extra=task_log_info)
else:
self.tasks[task_id]['status'] = States.done
self.tasks_completed_count += 1
logger.info("Task {} completed".format(task_id))
self.tasks[task_id]['time_returned'] = time.time()
if self.monitoring_config is not None:
task_log_info = self._create_task_log_info(task_id)
self.db_logger.info("Task Done", extra=task_log_info)
# it might be that in the course of the update, we've gone back to being
# pending - in which case, we should consider ourself for relaunch
if self.tasks[task_id]['status'] == States.pending:
self.launch_if_ready(task_id)
return
def handle_app_update(self, task_id, future, memo_cbk=False):
"""This function is called as a callback when an AppFuture
is in its final state.
It will trigger post-app processing such as checkpointing
and stageout.
Args:
task_id (string) : Task id
future (Future) : The relevant app future (which should be
consistent with the task structure 'app_fu' entry
KWargs:
memo_cbk(Bool) : Indicates that the call is coming from a memo update,
that does not require additional memo updates.
"""
if not self.tasks[task_id]['app_fu'].done():
logger.error(
"Internal consistency error: app_fu is not done for task {}".
format(task_id))
if not self.tasks[task_id]['app_fu'] == future:
logger.error(
"Internal consistency error: callback future is not the app_fu in task structure, for task {}"
.format(task_id))
if not memo_cbk:
# Update the memoizer with the new result if this is not a
# result from a memo lookup and the task has reached a terminal state.
self.memoizer.update_memo(task_id, self.tasks[task_id], future)
if self.checkpoint_mode == 'task_exit':
self.checkpoint(tasks=[task_id])
# Submit _*_stage_out tasks for output data futures that correspond with remote files
if (self.tasks[task_id]['app_fu']
and self.tasks[task_id]['app_fu'].done()
and self.tasks[task_id]['app_fu'].exception() is None
and self.tasks[task_id]['executor'] != 'data_manager'
and self.tasks[task_id]['func_name'] != '_file_stage_in'
and self.tasks[task_id]['func_name'] != '_ftp_stage_in'
and self.tasks[task_id]['func_name'] != '_http_stage_in'):
for dfu in self.tasks[task_id]['app_fu'].outputs:
f = dfu.file_obj
if isinstance(f, File) and f.is_remote():
f.stage_out(self.tasks[task_id]['executor'])
return
def launch_if_ready(self, task_id):
"""
launch_if_ready will launch the specified task, if it is ready
to run (for example, without dependencies, and in pending state).
This should be called by any piece of the DataFlowKernel that
thinks a task may have become ready to run.
It is not an error to call launch_if_ready on a task that is not
ready to run - launch_if_ready will not incorrectly launch that
task.
launch_if_ready is thread safe, so may be called from any thread
or callback.
"""
if self._count_deps(self.tasks[task_id]['depends']) == 0:
# We can now launch *task*
new_args, kwargs, exceptions = self.sanitize_and_wrap(
task_id, self.tasks[task_id]['args'],
self.tasks[task_id]['kwargs'])
self.tasks[task_id]['args'] = new_args
self.tasks[task_id]['kwargs'] = kwargs
if not exceptions:
# There are no dependency errors
exec_fu = None
# Acquire a lock, retest the state, launch
with self.tasks[task_id]['task_launch_lock']:
if self.tasks[task_id]['status'] == States.pending:
exec_fu = self.launch_task(task_id,
self.tasks[task_id]['func'],
*new_args, **kwargs)
if exec_fu:
self.tasks[task_id]['exec_fu'] = exec_fu
try:
self.tasks[task_id]['app_fu'].update_parent(exec_fu)
self.tasks[task_id]['exec_fu'] = exec_fu
except AttributeError as e:
logger.error(
"Task {}: Caught AttributeError at update_parent".
format(task_id))
raise e
else:
logger.info(
"Task {} failed due to dependency failure".format(task_id))
# Raise a dependency exception
self.tasks[task_id]['status'] = States.dep_fail
if self.monitoring_config is not None:
task_log_info = self._create_task_log_info(task_id, 'lazy')
self.db_logger.info("Task Dep Fail", extra=task_log_info)
try:
fu = Future()
fu.retries_left = 0
self.tasks[task_id]['exec_fu'] = fu
self.tasks[task_id]['app_fu'].update_parent(fu)
fu.set_exception(DependencyError(exceptions, task_id,
None))
except AttributeError as e:
logger.error(
"Task {} AttributeError at update_parent".format(
task_id))
raise e
def launch_task(self, task_id, executable, *args, **kwargs):
"""Handle the actual submission of the task to the executor layer.
If the app task has the executors attributes not set (default=='all')
the task is launched on a randomly selected executor from the
list of executors. This behavior could later be updated to support
binding to executors based on user specified criteria.
If the app task specifies a particular set of executors, it will be
targeted at those specific executors.
Args:
task_id (uuid string) : A uuid string that uniquely identifies the task
executable (callable) : A callable object
args (list of positional args)
kwargs (arbitrary keyword arguments)
Returns:
Future that tracks the execution of the submitted executable
"""
self.tasks[task_id]['time_submitted'] = time.time()
hit, memo_fu = self.memoizer.check_memo(task_id, self.tasks[task_id])
if hit:
logger.info("Reusing cached result for task {}".format(task_id))
try:
self.handle_exec_update(task_id, memo_fu)
except Exception as e:
logger.error(
"handle_exec_update raised an exception {} which will be ignored"
.format(e))
try:
self.handle_app_update(task_id, memo_fu, memo_cbk=True)
except Exception as e:
logger.error(
"handle_app_update raised an exception {} which will be ignored"
.format(e))
return memo_fu
executor_label = self.tasks[task_id]["executor"]
try:
executor = self.executors[executor_label]
except Exception:
logger.exception(
"Task {} requested invalid executor {}: config is\n{}".format(
task_id, executor_label, self._config))
if self.monitoring_config is not None:
executable = app_monitor.monitor_wrapper(executable, task_id,
self.monitoring_config,
self.run_id)
with self.submitter_lock:
exec_fu = executor.submit(executable, *args, **kwargs)
self.tasks[task_id]['status'] = States.running
if self.monitoring_config is not None:
task_log_info = self._create_task_log_info(task_id)
self.db_logger.info("Task Launch", extra=task_log_info)
exec_fu.retries_left = self._config.retries - \
self.tasks[task_id]['fail_count']
logger.info("Task {} launched on executor {}".format(
task_id, executor.label))
try:
exec_fu.add_done_callback(partial(self.handle_exec_update,
task_id))
except Exception as e:
logger.error(
"add_done_callback got an exception {} which will be ignored".
format(e))
return exec_fu
def _add_input_deps(self, executor, args, kwargs):
"""Look for inputs of the app that are remote files. Submit stage_in
apps for such files and replace the file objects in the inputs list with
corresponding DataFuture objects.
Args:
- executor (str) : executor where the app is going to be launched
- args (List) : Positional args to app function
- kwargs (Dict) : Kwargs to app function
"""
# Return if the task is _*_stage_in
if executor == 'data_manager':
return args, kwargs
inputs = kwargs.get('inputs', [])
for idx, f in enumerate(inputs):
if isinstance(f, File) and f.is_remote():
inputs[idx] = f.stage_in(executor)
for kwarg, f in kwargs.items():
if isinstance(f, File) and f.is_remote():
kwargs[kwarg] = f.stage_in(executor)
newargs = list(args)
for idx, f in enumerate(newargs):
if isinstance(f, File) and f.is_remote():
newargs[idx] = f.stage_in(executor)
return tuple(newargs), kwargs
def _gather_all_deps(self, args, kwargs):
"""Count the number of unresolved futures on which a task depends.
Args:
- args (List[args]) : The list of args list to the fn
- kwargs (Dict{kwargs}) : The dict of all kwargs passed to the fn
Returns:
- count, [list of dependencies]
"""
# Check the positional args
depends = []
count = 0
for dep in args:
if isinstance(dep, Future):
if self.tasks[dep.tid]['status'] not in FINAL_STATES:
count += 1
depends.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future):
if self.tasks[dep.tid]['status'] not in FINAL_STATES:
count += 1
depends.extend([dep])
# Check for futures in inputs=[<fut>...]
for dep in kwargs.get('inputs', []):
if isinstance(dep, Future):
if self.tasks[dep.tid]['status'] not in FINAL_STATES:
count += 1
depends.extend([dep])
return count, depends
def sanitize_and_wrap(self, task_id, args, kwargs):
"""This function should be called **ONLY** when all the futures we track have been resolved.
If the user hid futures a level below, we will not catch
it, and will (most likely) result in a type error.
Args:
task_id (uuid str) : Task id
func (Function) : App function
args (List) : Positional args to app function
kwargs (Dict) : Kwargs to app function
Return:
partial Function evaluated with all dependencies in args, kwargs and kwargs['inputs'] evaluated.
"""
dep_failures = []
# Replace item in args
new_args = []
for dep in args:
if isinstance(dep, Future):
try:
new_args.extend([dep.result()])
except Exception as e:
if self.tasks[dep.tid]['status'] in FINAL_FAILURE_STATES:
dep_failures.extend([e])
else:
new_args.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future):
try:
kwargs[key] = dep.result()
except Exception as e:
if self.tasks[dep.tid]['status'] in FINAL_FAILURE_STATES:
dep_failures.extend([e])
# Check for futures in inputs=[<fut>...]
if 'inputs' in kwargs:
new_inputs = []
for dep in kwargs['inputs']:
if isinstance(dep, Future):
try:
new_inputs.extend([dep.result()])
except Exception as e:
if self.tasks[
dep.tid]['status'] in FINAL_FAILURE_STATES:
dep_failures.extend([e])
else:
new_inputs.extend([dep])
kwargs['inputs'] = new_inputs
return new_args, kwargs, dep_failures
def submit(self,
func,
*args,
executors='all',
fn_hash=None,
cache=False,
**kwargs):
"""Add task to the dataflow system.
If the app task has the executors attributes not set (default=='all')
the task will be launched on a randomly selected executor from the
list of executors. If the app task specifies a particular set of
executors, it will be targeted at the specified executors.
>>> IF all deps are met:
>>> send to the runnable queue and launch the task
>>> ELSE:
>>> post the task in the pending queue
Args:
- func : A function object
- *args : Args to the function
KWargs :
- executors (list or string) : List of executors this call could go to.
Default='all'
- fn_hash (Str) : Hash of the function and inputs
Default=None
- cache (Bool) : To enable memoization or not
- kwargs (dict) : Rest of the kwargs to the fn passed as dict.
Returns:
(AppFuture) [DataFutures,]
"""
if self.cleanup_called:
raise ValueError("Cannot submit to a DFK that has been cleaned up")
task_id = self.task_count
self.task_count += 1
if isinstance(executors, str) and executors.lower() == 'all':
choices = list(e for e in self.executors if e != 'data_manager')
elif isinstance(executors, list):
choices = executors
executor = random.choice(choices)
# Transform remote input files to data futures
args, kwargs = self._add_input_deps(executor, args, kwargs)
task_def = {
'depends': None,
'executor': executor,
'func': func,
'func_name': func.__name__,
'args': args,
'kwargs': kwargs,
'fn_hash': fn_hash,
'memoize': cache,
'callback': None,
'exec_fu': None,
'checkpoint': None,
'fail_count': 0,
'fail_history': [],
'env': None,
'status': States.unsched,
'id': task_id,
'time_submitted': None,
'time_returned': None,
'app_fu': None
}
if task_id in self.tasks:
raise DuplicateTaskError(
"internal consistency error: Task {0} already exists in task list"
.format(task_id))
else:
self.tasks[task_id] = task_def
# Get the dep count and a list of dependencies for the task
dep_cnt, depends = self._gather_all_deps(args, kwargs)
self.tasks[task_id]['depends'] = depends
# Extract stdout and stderr to pass to AppFuture:
task_stdout = kwargs.get('stdout')
task_stderr = kwargs.get('stderr')
logger.info("Task {} submitted for App {}, waiting on tasks {}".format(
task_id, task_def['func_name'], [fu.tid for fu in depends]))
self.tasks[task_id]['task_launch_lock'] = threading.Lock()
app_fu = AppFuture(None,
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['app_fu'] = app_fu
app_fu.add_done_callback(partial(self.handle_app_update, task_id))
self.tasks[task_id]['status'] = States.pending
logger.debug("Task {} set to pending state with AppFuture: {}".format(
task_id, task_def['app_fu']))
# at this point add callbacks to all dependencies to do a launch_if_ready
# call whenever a dependency completes.
# we need to be careful about the order of setting the state to pending,
# adding the callbacks, and caling launch_if_ready explicitly once always below.
# I think as long as we call launch_if_ready once after setting pending, then
# we can add the callback dependencies at any point: if the callbacks all fire
# before then, they won't cause a launch, but the one below will. if they fire
# after we set it pending, then the last one will cause a launch, and the
# explicit one won't.
for d in depends:
def callback_adapter(dep_fut):
self.launch_if_ready(task_id)
try:
d.add_done_callback(callback_adapter)
except Exception as e:
logger.error(
"add_done_callback got an exception {} which will be ignored"
.format(e))
self.launch_if_ready(task_id)
return task_def['app_fu']
# it might also be interesting to assert that all DFK
# tasks are in a "final" state (3,4,5) when the DFK
# is closed down, and report some kind of warning.
# although really I'd like this to drain properly...
# and a drain function might look like this.
# If tasks have their states changed, this won't work properly
# but we can validate that...
def log_task_states(self):
logger.info("Summary of tasks in DFK:")
total_summarised = 0
keytasks = []
for tid in self.tasks:
keytasks.append((self.tasks[tid]['status'], tid))
def first(t):
return t[0]
sorted_keytasks = sorted(keytasks, key=first)
grouped_sorted_keytasks = itertools.groupby(sorted_keytasks, key=first)
# caution: g is an iterator that also advances the
# grouped_sorted_tasks iterator, so looping over
# both grouped_sorted_keytasks and g can only be done
# in certain patterns
for k, g in grouped_sorted_keytasks:
ts = []
for t in g:
tid = t[1]
ts.append(str(tid))
total_summarised = total_summarised + 1
tids_string = ", ".join(ts)
logger.info("Tasks in state {}: {}".format(str(k), tids_string))
total_in_tasks = len(self.tasks)
if total_summarised != total_in_tasks:
logger.error(
"Task count summarisation was inconsistent: summarised {} tasks, but tasks list contains {} tasks"
.format(total_summarised, total_in_tasks))
logger.info("End of summary")
def atexit_cleanup(self):
if not self.cleanup_called:
self.cleanup()
def wait_for_current_tasks(self):
"""Waits for all tasks in the task list to be completed, by waiting for their
AppFuture to be completed. This method will not necessarily wait for any tasks
added after cleanup has started (such as data stageout?)
"""
logger.info("Waiting for all remaining tasks to complete")
for task_id in self.tasks:
# .exception() is a less exception throwing way of
# waiting for completion than .result()
fut = self.tasks[task_id]['app_fu']
if not fut.done():
logger.debug("Waiting for task {} to complete".format(task_id))
fut.exception()
logger.info("All remaining tasks completed")
def cleanup(self):
"""DataFlowKernel cleanup.
This involves killing resources explicitly and sending die messages to IPP workers.
If the executors are managed (created by the DFK), then we call scale_in on each of
the executors and call executor.shutdown. Otherwise, we do nothing, and executor
cleanup is left to the user.
"""
logger.info("DFK cleanup initiated")
# this check won't detect two DFK cleanups happening from
# different threads extremely close in time because of
# non-atomic read/modify of self.cleanup_called
if self.cleanup_called:
raise Exception(
"attempt to clean up DFK when it has already been cleaned-up")
self.cleanup_called = True
self.log_task_states()
# Checkpointing takes priority over the rest of the tasks
# checkpoint if any valid checkpoint method is specified
if self.checkpoint_mode is not None:
self.checkpoint()
if self._checkpoint_timer:
logger.info("Stopping checkpoint timer")
self._checkpoint_timer.close()
# Send final stats
self.usage_tracker.send_message()
self.usage_tracker.close()
logger.info("Terminating flow_control and strategy threads")
self.flowcontrol.close()
for executor in self.executors.values():
if executor.managed:
if executor.scaling_enabled:
job_ids = executor.provider.resources.keys()
executor.scale_in(len(job_ids))
executor.shutdown()
self.time_completed = time.time()
self.db_logger.info("DFK end",
extra={
'tasks_failed_count': self.tasks_failed_count,
'tasks_completed_count':
self.tasks_completed_count,
"time_began": str(self.time_began),
'time_completed': str(self.time_completed),
'run_id': self.run_id,
'rundir': self.run_dir
})
if self.logging_server is not None:
self.logging_server.terminate()
self.logging_server.join()
if self.web_app is not None:
self.web_app.terminate()
self.web_app.join()
logger.info("DFK cleanup complete")
def checkpoint(self, tasks=None):
"""Checkpoint the dfk incrementally to a checkpoint file.
When called, every task that has been completed yet not
checkpointed is checkpointed to a file.
Kwargs:
- tasks (List of task ids) : List of task ids to checkpoint. Default=None
if set to None, we iterate over all tasks held by the DFK.
.. note::
Checkpointing only works if memoization is enabled
Returns:
Checkpoint dir if checkpoints were written successfully.
By default the checkpoints are written to the RUNDIR of the current
run under RUNDIR/checkpoints/{tasks.pkl, dfk.pkl}
"""
with self.checkpoint_lock:
checkpoint_queue = None
if tasks:
checkpoint_queue = tasks
else:
checkpoint_queue = self.tasks
checkpoint_dir = '{0}/checkpoint'.format(self.run_dir)
checkpoint_dfk = checkpoint_dir + '/dfk.pkl'
checkpoint_tasks = checkpoint_dir + '/tasks.pkl'
if not os.path.exists(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except FileExistsError:
pass
with open(checkpoint_dfk, 'wb') as f:
state = {'rundir': self.run_dir, 'task_count': self.task_count}
pickle.dump(state, f)
count = 0
with open(checkpoint_tasks, 'ab') as f:
for task_id in checkpoint_queue:
if not self.tasks[task_id]['checkpoint'] and \
self.tasks[task_id]['app_fu'].done() and \
self.tasks[task_id]['app_fu'].exception() is None:
hashsum = self.tasks[task_id]['hashsum']
if not hashsum:
continue
t = {
'hash': hashsum,
'exception': None,
'result': None
}
try:
# Asking for the result will raise an exception if
# the app had failed. Should we even checkpoint these?
# TODO : Resolve this question ?
r = self.memoizer.hash_lookup(hashsum).result()
except Exception as e:
t['exception'] = e
else:
t['result'] = r
# We are using pickle here since pickle dumps to a file in 'ab'
# mode behave like a incremental log.
pickle.dump(t, f)
count += 1
self.tasks[task_id]['checkpoint'] = True
logger.debug("Task {} checkpointed".format(task_id))
self.checkpointed_tasks += count
if count == 0:
if self.checkpointed_tasks == 0:
logger.warn(
"No tasks checkpointed so far in this run. Please ensure caching is enabled"
)
else:
logger.debug("No tasks checkpointed in this pass.")
else:
logger.info("Done checkpointing {} tasks".format(count))
return checkpoint_dir
def _load_checkpoints(self, checkpointDirs):
"""Load a checkpoint file into a lookup table.
The data being loaded from the pickle file mostly contains input
attributes of the task: func, args, kwargs, env...
To simplify the check of whether the exact task has been completed
in the checkpoint, we hash these input params and use it as the key
for the memoized lookup table.
Args:
- checkpointDirs (list) : List of filepaths to checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- memoized_lookup_table (dict)
"""
memo_lookup_table = {}
for checkpoint_dir in checkpointDirs:
logger.info("Loading checkpoints from {}".format(checkpoint_dir))
checkpoint_file = os.path.join(checkpoint_dir, 'tasks.pkl')
try:
with open(checkpoint_file, 'rb') as f:
while True:
try:
data = pickle.load(f)
# Copy and hash only the input attributes
memo_fu = Future()
if data['exception']:
memo_fu.set_exception(data['exception'])
else:
memo_fu.set_result(data['result'])
memo_lookup_table[data['hash']] = memo_fu
except EOFError:
# Done with the checkpoint file
break
except FileNotFoundError:
reason = "Checkpoint file was not found: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
except Exception:
reason = "Failed to load checkpoint: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
logger.info(
"Completed loading checkpoint:{0} with {1} tasks".format(
checkpoint_file, len(memo_lookup_table.keys())))
return memo_lookup_table
def load_checkpoints(self, checkpointDirs):
"""Load checkpoints from the checkpoint files into a dictionary.
The results are used to pre-populate the memoizer's lookup_table
Kwargs:
- checkpointDirs (list) : List of run folder to use as checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- dict containing, hashed -> future mappings
"""
self.memo_lookup_table = None
if not checkpointDirs:
return {}
if type(checkpointDirs) is not list:
raise BadCheckpoint("checkpointDirs expects a list of checkpoints")
return self._load_checkpoints(checkpointDirs)
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html')
self._config = config
self.run_dir = make_rundir(config.run_dir)
if config.initialize_logging:
parsl.set_file_logger("{}/parsl.log".format(self.run_dir), level=logging.DEBUG)
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
if sys.version_info < (3, 6):
logger.warning("Support for python versions < 3.6 is deprecated and will be removed after parsl 0.10")
logger.info("Parsl version: {}".format(get_version()))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# Monitoring
self.run_id = str(uuid4())
self.tasks_completed_count = 0
self.tasks_failed_count = 0
self.tasks_dep_fail_count = 0
self.monitoring = config.monitoring
# hub address and port for interchange to connect
self.hub_address = None
self.hub_interchange_port = None
if self.monitoring:
if self.monitoring.logdir is None:
self.monitoring.logdir = self.run_dir
self.hub_address = self.monitoring.hub_address
self.hub_interchange_port = self.monitoring.start(self.run_id)
self.time_began = datetime.datetime.now()
self.time_completed = None
# TODO: make configurable
logger.info("Run id is: " + self.run_id)
self.workflow_name = None
if self.monitoring is not None and self.monitoring.workflow_name is not None:
self.workflow_name = self.monitoring.workflow_name
else:
for frame in inspect.stack():
fname = os.path.basename(str(frame.filename))
parsl_file_names = ['dflow.py', 'typeguard.py']
# Find first file name not considered a parsl file
if fname not in parsl_file_names:
self.workflow_name = fname
break
self.workflow_version = str(self.time_began.replace(microsecond=0))
if self.monitoring is not None and self.monitoring.workflow_version is not None:
self.workflow_version = self.monitoring.workflow_version
workflow_info = {
'python_version': "{}.{}.{}".format(sys.version_info.major,
sys.version_info.minor,
sys.version_info.micro),
'parsl_version': get_version(),
"time_began": self.time_began,
'time_completed': None,
'workflow_duration': None,
'run_id': self.run_id,
'workflow_name': self.workflow_name,
'workflow_version': self.workflow_version,
'rundir': self.run_dir,
'tasks_completed_count': self.tasks_completed_count,
'tasks_failed_count': self.tasks_failed_count,
'user': getuser(),
'host': gethostname(),
}
if self.monitoring:
self.monitoring.send(MessageType.WORKFLOW_INFO,
workflow_info)
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self, memoize=config.app_cache, checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
self.data_manager = DataManager(self)
self.executors = {}
data_manager_executor = ThreadPoolExecutor(max_threads=config.data_management_max_threads, label='data_manager')
self.add_executors(config.executors + [data_manager_executor])
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint, interval=checkpoint_period, name="Checkpoint")
except Exception:
logger.error("invalid checkpoint_period provided: {0} expected HH:MM:SS".format(config.checkpoint_period))
self._checkpoint_timer = Timer(self.checkpoint, interval=(30 * 60), name="Checkpoint")
# if we use the functionality of dynamically adding executors
# all executors should be managed.
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.tasks = {}
self.submitter_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
class DataFlowKernel(object):
"""The DataFlowKernel adds dependency awareness to an existing executor.
It is responsible for managing futures, such that when dependencies are resolved,
pending tasks move to the runnable state.
Here is a simplified diagram of what happens internally::
User | DFK | Executor
----------------------------------------------------------
| |
Task-------+> +Submit |
App_Fu<------+--| |
| Dependencies met |
| task-------+--> +Submit
| Ex_Fu<------+----|
"""
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html')
self._config = config
self.run_dir = make_rundir(config.run_dir)
if config.initialize_logging:
parsl.set_file_logger("{}/parsl.log".format(self.run_dir), level=logging.DEBUG)
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
if sys.version_info < (3, 6):
logger.warning("Support for python versions < 3.6 is deprecated and will be removed after parsl 0.10")
logger.info("Parsl version: {}".format(get_version()))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# Monitoring
self.run_id = str(uuid4())
self.tasks_completed_count = 0
self.tasks_failed_count = 0
self.tasks_dep_fail_count = 0
self.monitoring = config.monitoring
# hub address and port for interchange to connect
self.hub_address = None
self.hub_interchange_port = None
if self.monitoring:
if self.monitoring.logdir is None:
self.monitoring.logdir = self.run_dir
self.hub_address = self.monitoring.hub_address
self.hub_interchange_port = self.monitoring.start(self.run_id)
self.time_began = datetime.datetime.now()
self.time_completed = None
# TODO: make configurable
logger.info("Run id is: " + self.run_id)
self.workflow_name = None
if self.monitoring is not None and self.monitoring.workflow_name is not None:
self.workflow_name = self.monitoring.workflow_name
else:
for frame in inspect.stack():
fname = os.path.basename(str(frame.filename))
parsl_file_names = ['dflow.py', 'typeguard.py']
# Find first file name not considered a parsl file
if fname not in parsl_file_names:
self.workflow_name = fname
break
self.workflow_version = str(self.time_began.replace(microsecond=0))
if self.monitoring is not None and self.monitoring.workflow_version is not None:
self.workflow_version = self.monitoring.workflow_version
workflow_info = {
'python_version': "{}.{}.{}".format(sys.version_info.major,
sys.version_info.minor,
sys.version_info.micro),
'parsl_version': get_version(),
"time_began": self.time_began,
'time_completed': None,
'workflow_duration': None,
'run_id': self.run_id,
'workflow_name': self.workflow_name,
'workflow_version': self.workflow_version,
'rundir': self.run_dir,
'tasks_completed_count': self.tasks_completed_count,
'tasks_failed_count': self.tasks_failed_count,
'user': getuser(),
'host': gethostname(),
}
if self.monitoring:
self.monitoring.send(MessageType.WORKFLOW_INFO,
workflow_info)
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self, memoize=config.app_cache, checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
self.data_manager = DataManager(self)
self.executors = {}
data_manager_executor = ThreadPoolExecutor(max_threads=config.data_management_max_threads, label='data_manager')
self.add_executors(config.executors + [data_manager_executor])
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint, interval=checkpoint_period, name="Checkpoint")
except Exception:
logger.error("invalid checkpoint_period provided: {0} expected HH:MM:SS".format(config.checkpoint_period))
self._checkpoint_timer = Timer(self.checkpoint, interval=(30 * 60), name="Checkpoint")
# if we use the functionality of dynamically adding executors
# all executors should be managed.
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.tasks = {}
self.submitter_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
def _create_task_log_info(self, task_id, fail_mode):
"""
Create the dictionary that will be included in the log.
"""
info_to_monitor = ['func_name', 'fn_hash', 'memoize', 'fail_count', 'status',
'id', 'time_submitted', 'time_returned', 'executor']
task_log_info = {"task_" + k: self.tasks[task_id][k] for k in info_to_monitor}
task_log_info['run_id'] = self.run_id
task_log_info['timestamp'] = datetime.datetime.now()
task_log_info['task_status_name'] = self.tasks[task_id]['status'].name
task_log_info['tasks_failed_count'] = self.tasks_failed_count
task_log_info['tasks_completed_count'] = self.tasks_completed_count
task_log_info['task_inputs'] = str(self.tasks[task_id]['kwargs'].get('inputs', None))
task_log_info['task_outputs'] = str(self.tasks[task_id]['kwargs'].get('outputs', None))
task_log_info['task_stdin'] = self.tasks[task_id]['kwargs'].get('stdin', None)
stdout_spec = self.tasks[task_id]['kwargs'].get('stdout', None)
stderr_spec = self.tasks[task_id]['kwargs'].get('stderr', None)
try:
stdout_name, stdout_mode = get_std_fname_mode('stdout', stdout_spec)
except Exception as e:
logger.warning("Incorrect stdout format {} for Task {}".format(stdout_spec, task_id))
stdout_name, stdout_mode = str(e), None
try:
stderr_name, stderr_mode = get_std_fname_mode('stderr', stderr_spec)
except Exception as e:
logger.warning("Incorrect stderr format {} for Task {}".format(stderr_spec, task_id))
stderr_name, stderr_mode = str(e), None
stdout_spec = ";".join((stdout_name, stdout_mode)) if stdout_mode else stdout_name
stderr_spec = ";".join((stderr_name, stderr_mode)) if stderr_mode else stderr_name
task_log_info['task_stdout'] = stdout_spec
task_log_info['task_stderr'] = stderr_spec
task_log_info['task_fail_history'] = None
if self.tasks[task_id]['fail_history'] is not None:
task_log_info['task_fail_history'] = ",".join(self.tasks[task_id]['fail_history'])
task_log_info['task_depends'] = None
if self.tasks[task_id]['depends'] is not None:
task_log_info['task_depends'] = ",".join([str(t.tid) for t in self.tasks[task_id]['depends']
if isinstance(t, AppFuture) or isinstance(t, DataFuture)])
task_log_info['task_elapsed_time'] = None
if self.tasks[task_id]['time_returned'] is not None:
task_log_info['task_elapsed_time'] = (self.tasks[task_id]['time_returned'] -
self.tasks[task_id]['time_submitted']).total_seconds()
task_log_info['task_fail_mode'] = fail_mode
return task_log_info
def _count_deps(self, depends):
"""Internal.
Count the number of unresolved futures in the list depends.
"""
count = 0
for dep in depends:
if isinstance(dep, Future):
if not dep.done():
count += 1
return count
@property
def config(self):
"""Returns the fully initialized config that the DFK is actively using.
Returns:
- config (dict)
"""
return self._config
def handle_exec_update(self, task_id, future):
"""This function is called only as a callback from an execution
attempt reaching a final state (either successfully or failing).
It will launch retries if necessary, and update the task
structure.
Args:
task_id (string) : Task id which is a uuid string
future (Future) : The future object corresponding to the task which
makes this callback
"""
try:
res = future.result()
if isinstance(res, RemoteExceptionWrapper):
res.reraise()
except Exception as e:
logger.debug("Task {} failed".format(task_id))
# We keep the history separately, since the future itself could be
# tossed.
self.tasks[task_id]['fail_history'].append(str(e))
self.tasks[task_id]['fail_count'] += 1
if not self._config.lazy_errors:
logger.exception("Eager fail, skipping retry logic")
self.tasks[task_id]['status'] = States.failed
if self.monitoring:
task_log_info = self._create_task_log_info(task_id, 'eager')
self.monitoring.send(MessageType.TASK_INFO, task_log_info)
return
if self.tasks[task_id]['status'] == States.dep_fail:
logger.info("Task {} failed due to dependency failure so skipping retries".format(task_id))
elif self.tasks[task_id]['fail_count'] <= self._config.retries:
self.tasks[task_id]['status'] = States.pending
logger.info("Task {} marked for retry".format(task_id))
else:
logger.exception("Task {} failed after {} retry attempts".format(task_id,
self._config.retries))
self.tasks[task_id]['status'] = States.failed
self.tasks_failed_count += 1
self.tasks[task_id]['time_returned'] = datetime.datetime.now()
else:
self.tasks[task_id]['status'] = States.done
self.tasks_completed_count += 1
logger.info("Task {} completed".format(task_id))
self.tasks[task_id]['time_returned'] = datetime.datetime.now()
if self.tasks[task_id]['app_fu'].stdout is not None:
logger.info("Standard output for task {} available at {}".format(task_id, self.tasks[task_id]['app_fu'].stdout))
if self.tasks[task_id]['app_fu'].stderr is not None:
logger.info("Standard error for task {} available at {}".format(task_id, self.tasks[task_id]['app_fu'].stderr))
if self.monitoring:
task_log_info = self._create_task_log_info(task_id, 'lazy')
self.monitoring.send(MessageType.TASK_INFO, task_log_info)
# it might be that in the course of the update, we've gone back to being
# pending - in which case, we should consider ourself for relaunch
if self.tasks[task_id]['status'] == States.pending:
self.launch_if_ready(task_id)
with self.tasks[task_id]['app_fu']._update_lock:
if not future.done():
raise ValueError("done callback called, despite future not reporting itself as done")
try:
res = future.result()
if isinstance(res, RemoteExceptionWrapper):
res.reraise()
self.tasks[task_id]['app_fu'].set_result(future.result())
except Exception as e:
if future.retries_left > 0:
# ignore this exception, because assume some later
# parent executor, started external to this class,
# will provide the answer
pass
else:
self.tasks[task_id]['app_fu'].set_exception(e)
return
def handle_app_update(self, task_id, future, memo_cbk=False):
"""This function is called as a callback when an AppFuture
is in its final state.
It will trigger post-app processing such as checkpointing.
Args:
task_id (string) : Task id
future (Future) : The relevant app future (which should be
consistent with the task structure 'app_fu' entry
KWargs:
memo_cbk(Bool) : Indicates that the call is coming from a memo update,
that does not require additional memo updates.
"""
if not self.tasks[task_id]['app_fu'].done():
logger.error("Internal consistency error: app_fu is not done for task {}".format(task_id))
if not self.tasks[task_id]['app_fu'] == future:
logger.error("Internal consistency error: callback future is not the app_fu in task structure, for task {}".format(task_id))
if not memo_cbk:
# Update the memoizer with the new result if this is not a
# result from a memo lookup and the task has reached a terminal state.
self.memoizer.update_memo(task_id, self.tasks[task_id], future)
if self.checkpoint_mode == 'task_exit':
self.checkpoint(tasks=[task_id])
# If checkpointing is turned on, wiping app_fu is left to the checkpointing code
# else we wipe it here.
if self.checkpoint_mode is None:
self.wipe_task(task_id)
return
def wipe_task(self, task_id):
""" Remove task with task_id from the internal tasks table
"""
del self.tasks[task_id]
@staticmethod
def check_staging_inhibited(kwargs):
return kwargs.get('staging_inhibit_output', False)
def launch_if_ready(self, task_id):
"""
launch_if_ready will launch the specified task, if it is ready
to run (for example, without dependencies, and in pending state).
This should be called by any piece of the DataFlowKernel that
thinks a task may have become ready to run.
It is not an error to call launch_if_ready on a task that is not
ready to run - launch_if_ready will not incorrectly launch that
task.
launch_if_ready is thread safe, so may be called from any thread
or callback.
"""
# after launching the task, self.tasks[task_id] is no longer
# guaranteed to exist (because it can complete fast as part of the
# submission - eg memoization)
task_record = self.tasks.get(task_id)
if task_record is None:
# assume this task has already been processed to completion
logger.info("Task {} has no task record. Assuming it has already been processed to completion.".format(task_id))
return
if self._count_deps(task_record['depends']) == 0:
# We can now launch *task*
new_args, kwargs, exceptions = self.sanitize_and_wrap(task_id,
task_record['args'],
task_record['kwargs'])
task_record['args'] = new_args
task_record['kwargs'] = kwargs
if not exceptions:
# There are no dependency errors
exec_fu = None
# Acquire a lock, retest the state, launch
with task_record['task_launch_lock']:
if task_record['status'] == States.pending:
exec_fu = self.launch_task(
task_id, task_record['func'], *new_args, **kwargs)
else:
logger.info(
"Task {} failed due to dependency failure".format(task_id))
# Raise a dependency exception
task_record['status'] = States.dep_fail
self.tasks_dep_fail_count += 1
if self.monitoring is not None:
task_log_info = self._create_task_log_info(task_id, 'lazy')
self.monitoring.send(MessageType.TASK_INFO, task_log_info)
exec_fu = Future()
exec_fu.retries_left = 0
exec_fu.set_exception(DependencyError(exceptions,
task_id,
None))
if exec_fu:
try:
exec_fu.add_done_callback(partial(self.handle_exec_update, task_id))
except Exception as e:
logger.error("add_done_callback got an exception {} which will be ignored".format(e))
task_record['exec_fu'] = exec_fu
def launch_task(self, task_id, executable, *args, **kwargs):
"""Handle the actual submission of the task to the executor layer.
If the app task has the executors attributes not set (default=='all')
the task is launched on a randomly selected executor from the
list of executors. This behavior could later be updated to support
binding to executors based on user specified criteria.
If the app task specifies a particular set of executors, it will be
targeted at those specific executors.
Args:
task_id (uuid string) : A uuid string that uniquely identifies the task
executable (callable) : A callable object
args (list of positional args)
kwargs (arbitrary keyword arguments)
Returns:
Future that tracks the execution of the submitted executable
"""
self.tasks[task_id]['time_submitted'] = datetime.datetime.now()
hit, memo_fu = self.memoizer.check_memo(task_id, self.tasks[task_id])
if hit:
logger.info("Reusing cached result for task {}".format(task_id))
return memo_fu
executor_label = self.tasks[task_id]["executor"]
try:
executor = self.executors[executor_label]
except Exception:
logger.exception("Task {} requested invalid executor {}: config is\n{}".format(task_id, executor_label, self._config))
raise ValueError("Task {} requested invalid executor {}".format(task_id, executor_label))
if self.monitoring is not None and self.monitoring.resource_monitoring_enabled:
wrapper_logging_level = logging.DEBUG if self.monitoring.monitoring_debug else logging.INFO
executable = self.monitoring.monitor_wrapper(executable, task_id,
self.monitoring.monitoring_hub_url,
self.run_id,
wrapper_logging_level,
self.monitoring.resource_monitoring_interval)
with self.submitter_lock:
exec_fu = executor.submit(executable, *args, **kwargs)
self.tasks[task_id]['status'] = States.launched
if self.monitoring is not None:
task_log_info = self._create_task_log_info(task_id, 'lazy')
self.monitoring.send(MessageType.TASK_INFO, task_log_info)
exec_fu.retries_left = self._config.retries - \
self.tasks[task_id]['fail_count']
logger.info("Task {} launched on executor {}".format(task_id, executor.label))
return exec_fu
def _add_input_deps(self, executor, args, kwargs, func):
"""Look for inputs of the app that are files. Give the data manager
the opportunity to replace a file with a data future for that file,
for example wrapping the result of a staging action.
Args:
- executor (str) : executor where the app is going to be launched
- args (List) : Positional args to app function
- kwargs (Dict) : Kwargs to app function
"""
# Return if the task is _*_stage_in
if executor == 'data_manager':
return args, kwargs, func
inputs = kwargs.get('inputs', [])
for idx, f in enumerate(inputs):
(inputs[idx], func) = self.data_manager.optionally_stage_in(f, func, executor)
for kwarg, f in kwargs.items():
(kwargs[kwarg], func) = self.data_manager.optionally_stage_in(f, func, executor)
newargs = list(args)
for idx, f in enumerate(newargs):
(newargs[idx], func) = self.data_manager.optionally_stage_in(f, func, executor)
return tuple(newargs), kwargs, func
def _add_output_deps(self, executor, args, kwargs, app_fut, func):
logger.debug("Adding output dependencies")
outputs = kwargs.get('outputs', [])
app_fut._outputs = []
for idx, f in enumerate(outputs):
if isinstance(f, File) and not self.check_staging_inhibited(kwargs):
# replace a File with a DataFuture - either completing when the stageout
# future completes, or if no stage out future is returned, then when the
# app itself completes.
# The staging code will get a clean copy which it is allowed to mutate,
# while the DataFuture-contained original will not be modified by any staging.
f_copy = f.cleancopy()
outputs[idx] = f_copy
logger.debug("Submitting stage out for output file {}".format(repr(f)))
stageout_fut = self.data_manager.stage_out(f_copy, executor, app_fut)
if stageout_fut:
logger.debug("Adding a dependency on stageout future for {}".format(repr(f)))
app_fut._outputs.append(DataFuture(stageout_fut, f, tid=app_fut.tid))
else:
logger.debug("No stageout dependency for {}".format(repr(f)))
app_fut._outputs.append(DataFuture(app_fut, f, tid=app_fut.tid))
# this is a hook for post-task stageout
# note that nothing depends on the output - which is maybe a bug
# in the not-very-tested stageout system?
newfunc = self.data_manager.replace_task_stage_out(f_copy, func, executor)
if newfunc:
func = newfunc
else:
logger.debug("Not performing staging for: {}".format(repr(f)))
app_fut._outputs.append(DataFuture(app_fut, f, tid=app_fut.tid))
return func
def _gather_all_deps(self, args, kwargs):
"""Count the number of unresolved futures on which a task depends.
Args:
- args (List[args]) : The list of args list to the fn
- kwargs (Dict{kwargs}) : The dict of all kwargs passed to the fn
Returns:
- count, [list of dependencies]
"""
# Check the positional args
depends = []
def check_dep(d):
if isinstance(d, Future):
depends.extend([d])
for dep in args:
check_dep(dep)
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
check_dep(dep)
# Check for futures in inputs=[<fut>...]
for dep in kwargs.get('inputs', []):
check_dep(dep)
return depends
def sanitize_and_wrap(self, task_id, args, kwargs):
"""This function should be called only when all the futures we track have been resolved.
If the user hid futures a level below, we will not catch
it, and will (most likely) result in a type error.
Args:
task_id (uuid str) : Task id
func (Function) : App function
args (List) : Positional args to app function
kwargs (Dict) : Kwargs to app function
Return:
partial function evaluated with all dependencies in args, kwargs and kwargs['inputs'] evaluated.
"""
dep_failures = []
# Replace item in args
new_args = []
for dep in args:
if isinstance(dep, Future):
try:
new_args.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_args.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future):
try:
kwargs[key] = dep.result()
except Exception as e:
dep_failures.extend([e])
# Check for futures in inputs=[<fut>...]
if 'inputs' in kwargs:
new_inputs = []
for dep in kwargs['inputs']:
if isinstance(dep, Future):
try:
new_inputs.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_inputs.extend([dep])
kwargs['inputs'] = new_inputs
return new_args, kwargs, dep_failures
def submit(self, func, *args, executors='all', fn_hash=None, cache=False, **kwargs):
"""Add task to the dataflow system.
If the app task has the executors attributes not set (default=='all')
the task will be launched on a randomly selected executor from the
list of executors. If the app task specifies a particular set of
executors, it will be targeted at the specified executors.
>>> IF all deps are met:
>>> send to the runnable queue and launch the task
>>> ELSE:
>>> post the task in the pending queue
Args:
- func : A function object
- *args : Args to the function
KWargs :
- executors (list or string) : List of executors this call could go to.
Default='all'
- fn_hash (Str) : Hash of the function and inputs
Default=None
- cache (Bool) : To enable memoization or not
- kwargs (dict) : Rest of the kwargs to the fn passed as dict.
Returns:
(AppFuture) [DataFutures,]
"""
if self.cleanup_called:
raise ValueError("Cannot submit to a DFK that has been cleaned up")
task_id = self.task_count
self.task_count += 1
if isinstance(executors, str) and executors.lower() == 'all':
choices = list(e for e in self.executors if e != 'data_manager')
elif isinstance(executors, list):
choices = executors
else:
raise ValueError("Task {} supplied invalid type for executors: {}".format(task_id, type(executors)))
executor = random.choice(choices)
# The below uses func.__name__ before it has been wrapped by any staging code.
label = kwargs.get('label')
for kw in ['stdout', 'stderr']:
if kw in kwargs:
if kwargs[kw] == parsl.AUTO_LOGNAME:
kwargs[kw] = os.path.join(
self.run_dir,
'task_logs',
str(int(task_id / 10000)).zfill(4), # limit logs to 10k entries per directory
'task_{}_{}{}.{}'.format(
str(task_id).zfill(4),
func.__name__,
'' if label is None else '_{}'.format(label),
kw)
)
task_def = {'depends': None,
'executor': executor,
'func_name': func.__name__,
'fn_hash': fn_hash,
'memoize': cache,
'exec_fu': None,
'fail_count': 0,
'fail_history': [],
'status': States.unsched,
'id': task_id,
'time_submitted': None,
'time_returned': None}
app_fu = AppFuture(task_def)
# Transform remote input files to data futures
args, kwargs, func = self._add_input_deps(executor, args, kwargs, func)
func = self._add_output_deps(executor, args, kwargs, app_fu, func)
task_def.update({
'args': args,
'func': func,
'kwargs': kwargs,
'app_fu': app_fu})
if task_id in self.tasks:
raise DuplicateTaskError(
"internal consistency error: Task {0} already exists in task list".format(task_id))
else:
self.tasks[task_id] = task_def
# Get the list of dependencies for the task
depends = self._gather_all_deps(args, kwargs)
self.tasks[task_id]['depends'] = depends
depend_descs = []
for d in depends:
if isinstance(d, AppFuture) or isinstance(d, DataFuture):
depend_descs.append("task {}".format(d.tid))
else:
depend_descs.append(repr(d))
if depend_descs != []:
waiting_message = "waiting on {}".format(", ".join(depend_descs))
else:
waiting_message = "not waiting on any dependency"
logger.info("Task {} submitted for App {}, {}".format(task_id,
task_def['func_name'],
waiting_message))
self.tasks[task_id]['task_launch_lock'] = threading.Lock()
app_fu.add_done_callback(partial(self.handle_app_update, task_id))
self.tasks[task_id]['status'] = States.pending
logger.debug("Task {} set to pending state with AppFuture: {}".format(task_id, task_def['app_fu']))
# at this point add callbacks to all dependencies to do a launch_if_ready
# call whenever a dependency completes.
# we need to be careful about the order of setting the state to pending,
# adding the callbacks, and caling launch_if_ready explicitly once always below.
# I think as long as we call launch_if_ready once after setting pending, then
# we can add the callback dependencies at any point: if the callbacks all fire
# before then, they won't cause a launch, but the one below will. if they fire
# after we set it pending, then the last one will cause a launch, and the
# explicit one won't.
for d in depends:
def callback_adapter(dep_fut):
self.launch_if_ready(task_id)
try:
d.add_done_callback(callback_adapter)
except Exception as e:
logger.error("add_done_callback got an exception {} which will be ignored".format(e))
self.launch_if_ready(task_id)
return app_fu
# it might also be interesting to assert that all DFK
# tasks are in a "final" state (3,4,5) when the DFK
# is closed down, and report some kind of warning.
# although really I'd like this to drain properly...
# and a drain function might look like this.
# If tasks have their states changed, this won't work properly
# but we can validate that...
def log_task_states(self):
logger.info("Summary of tasks in DFK:")
keytasks = {state: 0 for state in States}
for tid in self.tasks:
keytasks[self.tasks[tid]['status']] += 1
# Fetch from counters since tasks get wiped
keytasks[States.done] = self.tasks_completed_count
keytasks[States.failed] = self.tasks_failed_count
keytasks[States.dep_fail] = self.tasks_dep_fail_count
for state in States:
if keytasks[state]:
logger.info("Tasks in state {}: {}".format(str(state), keytasks[state]))
total_summarized = sum(keytasks.values())
if total_summarized != self.task_count:
logger.error("Task count summarisation was inconsistent: summarised {} tasks, but task counters registered {} tasks".format(
total_summarized, self.task_count))
logger.info("End of summary")
def _create_remote_dirs_over_channel(self, provider, channel):
""" Create script directories across a channel
Parameters
----------
provider: Provider obj
Provider for which scritps dirs are being created
channel: Channel obk
Channel over which the remote dirs are to be created
"""
run_dir = self.run_dir
if channel.script_dir is None:
channel.script_dir = os.path.join(run_dir, 'submit_scripts')
# Only create dirs if we aren't on a shared-fs
if not channel.isdir(run_dir):
parent, child = pathlib.Path(run_dir).parts[-2:]
remote_run_dir = os.path.join(parent, child)
channel.script_dir = os.path.join(remote_run_dir, 'remote_submit_scripts')
provider.script_dir = os.path.join(run_dir, 'local_submit_scripts')
channel.makedirs(channel.script_dir, exist_ok=True)
def add_executors(self, executors):
for executor in executors:
executor.run_dir = self.run_dir
executor.hub_address = self.hub_address
executor.hub_port = self.hub_interchange_port
if hasattr(executor, 'provider'):
if hasattr(executor.provider, 'script_dir'):
executor.provider.script_dir = os.path.join(self.run_dir, 'submit_scripts')
os.makedirs(executor.provider.script_dir, exist_ok=True)
if hasattr(executor.provider, 'channels'):
logger.debug("Creating script_dir across multiple channels")
for channel in executor.provider.channels:
self._create_remote_dirs_over_channel(executor.provider, channel)
else:
self._create_remote_dirs_over_channel(executor.provider, executor.provider.channel)
self.executors[executor.label] = executor
executor.start()
if hasattr(self, 'flowcontrol') and isinstance(self.flowcontrol, FlowControl):
self.flowcontrol.strategy.add_executors(executors)
def atexit_cleanup(self):
if not self.cleanup_called:
self.cleanup()
def wait_for_current_tasks(self):
"""Waits for all tasks in the task list to be completed, by waiting for their
AppFuture to be completed. This method will not necessarily wait for any tasks
added after cleanup has started (such as data stageout?)
"""
logger.info("Waiting for all remaining tasks to complete")
for task_id in self.tasks:
# .exception() is a less exception throwing way of
# waiting for completion than .result()
fut = self.tasks[task_id]['app_fu']
if not fut.done():
logger.debug("Waiting for task {} to complete".format(task_id))
fut.exception()
logger.info("All remaining tasks completed")
def cleanup(self):
"""DataFlowKernel cleanup.
This involves releasing all resources explicitly.
If the executors are managed by the DFK, then we call scale_in on each of
the executors and call executor.shutdown. Otherwise, executor cleanup is left to
the user.
"""
logger.info("DFK cleanup initiated")
# this check won't detect two DFK cleanups happening from
# different threads extremely close in time because of
# non-atomic read/modify of self.cleanup_called
if self.cleanup_called:
raise Exception("attempt to clean up DFK when it has already been cleaned-up")
self.cleanup_called = True
self.log_task_states()
# Checkpointing takes priority over the rest of the tasks
# checkpoint if any valid checkpoint method is specified
if self.checkpoint_mode is not None:
self.checkpoint()
if self._checkpoint_timer:
logger.info("Stopping checkpoint timer")
self._checkpoint_timer.close()
# Send final stats
self.usage_tracker.send_message()
self.usage_tracker.close()
logger.info("Terminating flow_control and strategy threads")
self.flowcontrol.close()
for executor in self.executors.values():
if executor.managed:
if executor.scaling_enabled:
job_ids = executor.provider.resources.keys()
executor.scale_in(len(job_ids))
executor.shutdown()
self.time_completed = datetime.datetime.now()
if self.monitoring:
self.monitoring.send(MessageType.WORKFLOW_INFO,
{'tasks_failed_count': self.tasks_failed_count,
'tasks_completed_count': self.tasks_completed_count,
"time_began": self.time_began,
'time_completed': self.time_completed,
'workflow_duration': (self.time_completed - self.time_began).total_seconds(),
'run_id': self.run_id, 'rundir': self.run_dir})
self.monitoring.close()
logger.info("DFK cleanup complete")
def checkpoint(self, tasks=None):
"""Checkpoint the dfk incrementally to a checkpoint file.
When called, every task that has been completed yet not
checkpointed is checkpointed to a file.
Kwargs:
- tasks (List of task ids) : List of task ids to checkpoint. Default=None
if set to None, we iterate over all tasks held by the DFK.
.. note::
Checkpointing only works if memoization is enabled
Returns:
Checkpoint dir if checkpoints were written successfully.
By default the checkpoints are written to the RUNDIR of the current
run under RUNDIR/checkpoints/{tasks.pkl, dfk.pkl}
"""
with self.checkpoint_lock:
checkpoint_queue = None
if tasks:
checkpoint_queue = tasks
else:
checkpoint_queue = list(self.tasks.keys())
checkpoint_dir = '{0}/checkpoint'.format(self.run_dir)
checkpoint_dfk = checkpoint_dir + '/dfk.pkl'
checkpoint_tasks = checkpoint_dir + '/tasks.pkl'
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir, exist_ok=True)
with open(checkpoint_dfk, 'wb') as f:
state = {'rundir': self.run_dir,
'task_count': self.task_count
}
pickle.dump(state, f)
count = 0
with open(checkpoint_tasks, 'ab') as f:
for task_id in checkpoint_queue:
if task_id in self.tasks and \
self.tasks[task_id]['app_fu'] is not None and \
self.tasks[task_id]['app_fu'].done() and \
self.tasks[task_id]['app_fu'].exception() is None:
hashsum = self.tasks[task_id]['hashsum']
self.wipe_task(task_id)
# self.tasks[task_id]['app_fu'] = None
if not hashsum:
continue
t = {'hash': hashsum,
'exception': None,
'result': None}
try:
# Asking for the result will raise an exception if
# the app had failed. Should we even checkpoint these?
# TODO : Resolve this question ?
r = self.memoizer.hash_lookup(hashsum).result()
except Exception as e:
t['exception'] = e
else:
t['result'] = r
# We are using pickle here since pickle dumps to a file in 'ab'
# mode behave like a incremental log.
pickle.dump(t, f)
count += 1
logger.debug("Task {} checkpointed".format(task_id))
self.checkpointed_tasks += count
if count == 0:
if self.checkpointed_tasks == 0:
logger.warning("No tasks checkpointed so far in this run. Please ensure caching is enabled")
else:
logger.debug("No tasks checkpointed in this pass.")
else:
logger.info("Done checkpointing {} tasks".format(count))
return checkpoint_dir
def _load_checkpoints(self, checkpointDirs):
"""Load a checkpoint file into a lookup table.
The data being loaded from the pickle file mostly contains input
attributes of the task: func, args, kwargs, env...
To simplify the check of whether the exact task has been completed
in the checkpoint, we hash these input params and use it as the key
for the memoized lookup table.
Args:
- checkpointDirs (list) : List of filepaths to checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- memoized_lookup_table (dict)
"""
memo_lookup_table = {}
for checkpoint_dir in checkpointDirs:
logger.info("Loading checkpoints from {}".format(checkpoint_dir))
checkpoint_file = os.path.join(checkpoint_dir, 'tasks.pkl')
try:
with open(checkpoint_file, 'rb') as f:
while True:
try:
data = pickle.load(f)
# Copy and hash only the input attributes
memo_fu = Future()
if data['exception']:
memo_fu.set_exception(data['exception'])
else:
memo_fu.set_result(data['result'])
memo_lookup_table[data['hash']] = memo_fu
except EOFError:
# Done with the checkpoint file
break
except FileNotFoundError:
reason = "Checkpoint file was not found: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
except Exception:
reason = "Failed to load checkpoint: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
logger.info("Completed loading checkpoint: {0} with {1} tasks".format(checkpoint_file,
len(memo_lookup_table.keys())))
return memo_lookup_table
def load_checkpoints(self, checkpointDirs):
"""Load checkpoints from the checkpoint files into a dictionary.
The results are used to pre-populate the memoizer's lookup_table
Kwargs:
- checkpointDirs (list) : List of run folder to use as checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- dict containing, hashed -> future mappings
"""
self.memo_lookup_table = None
if not checkpointDirs:
return {}
if type(checkpointDirs) is not list:
raise BadCheckpoint("checkpointDirs expects a list of checkpoints")
return self._load_checkpoints(checkpointDirs)
def __init__(self,
config=None,
executors=None,
lazyErrors=True,
appCache=True,
rundir=None,
retries=0,
checkpointFiles=None,
checkpointMode=None,
data_manager=None):
""" Initialize the DataFlowKernel.
Please note that keyword args passed to the DFK here will always override
options passed in via the config.
KWargs:
- config (dict) : A single data object encapsulating all config attributes
- executors (list of Executor objs): Optional, kept for (somewhat) backward compatibility with 0.2.0
- lazyErrors(bool) : Default=True, allow workflow to continue on app failures.
- appCache (bool) :Enable caching of apps
- rundir (str) : Path to run directory. Defaults to ./runinfo/runNNN
- retries(int): Default=0, Set the number of retry attempts in case of failure
- checkpointFiles (list of str): List of filepaths to checkpoint files
- checkpointMode (None, 'dfk_exit', 'task_exit', 'periodic'): Method to use.
- data_manager (DataManager): User created DataManager
Returns:
DataFlowKernel object
"""
# Create run dirs for this run
self.rundir = make_rundir(config=config, path=rundir)
parsl.set_file_logger("{}/parsl.log".format(self.rundir),
level=logging.DEBUG)
logger.info("Parsl version: {}".format(parsl.__version__))
logger.info("Libsubmit version: {}".format(libsubmit.__version__))
# Update config with defaults
self._config = update_config(config, self.rundir)
# Set the data manager
if data_manager:
self.data_manager = data_manager
else:
self.data_manager = DataManager(config=self._config)
# Start the anonymized usage tracker and send init msg
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# Load checkpoints if any
cpts = self.load_checkpoints(checkpointFiles)
# Initialize the memoizer
self.memoizer = Memoizer(self, memoize=appCache, checkpoint=cpts)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
if self._config:
self._executors_managed = True
# Create the executors
epf = EPF()
self.executors = epf.make(self.rundir, self._config)
# set global vars from config
self.lazy_fail = self._config["globals"].get(
"lazyErrors", lazyErrors)
self.fail_retries = self._config["globals"].get("retries", retries)
self.flowcontrol = FlowControl(self, self._config)
self.checkpoint_mode = self._config["globals"].get(
"checkpointMode", checkpointMode)
if self.checkpoint_mode == "periodic":
period = self._config["globals"].get("checkpointPeriod",
"00:30:00")
try:
h, m, s = map(int, period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception as e:
logger.error(
"invalid checkpointPeriod provided:{0} expected HH:MM:SS"
.format(period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
else:
self._executors_managed = False
self.fail_retries = retries
self.lazy_fail = lazyErrors
self.executors = {i: x for i, x in enumerate(executors)}
self.flowcontrol = FlowNoControl(self, None)
self.checkpoint_mode = checkpointMode
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
self.task_launch_lock = threading.Lock()
logger.debug("Using executors: {0}".format(self.executors))
atexit.register(self.cleanup)
class DataFlowKernel(object):
"""The DataFlowKernel adds dependency awareness to an existing executor.
It is responsible for managing futures, such that when dependencies are resolved,
pending tasks move to the runnable state.
Here's a simplified diagram of what happens internally::
User | DFK | Executor
----------------------------------------------------------
| |
Task-------+> +Submit |
App_Fu<------+--| |
| Dependencies met |
| task-------+--> +Submit
| Ex_Fu<------+----|
"""
def __init__(self,
config=None,
executors=None,
lazyErrors=True,
appCache=True,
rundir=None,
retries=0,
checkpointFiles=None,
checkpointMode=None,
data_manager=None):
""" Initialize the DataFlowKernel.
Please note that keyword args passed to the DFK here will always override
options passed in via the config.
KWargs:
- config (dict) : A single data object encapsulating all config attributes
- executors (list of Executor objs): Optional, kept for (somewhat) backward compatibility with 0.2.0
- lazyErrors(bool) : Default=True, allow workflow to continue on app failures.
- appCache (bool) :Enable caching of apps
- rundir (str) : Path to run directory. Defaults to ./runinfo/runNNN
- retries(int): Default=0, Set the number of retry attempts in case of failure
- checkpointFiles (list of str): List of filepaths to checkpoint files
- checkpointMode (None, 'dfk_exit', 'task_exit', 'periodic'): Method to use.
- data_manager (DataManager): User created DataManager
Returns:
DataFlowKernel object
"""
# Create run dirs for this run
self.rundir = make_rundir(config=config, path=rundir)
parsl.set_file_logger("{}/parsl.log".format(self.rundir),
level=logging.DEBUG)
logger.info("Parsl version: {}".format(parsl.__version__))
logger.info("Libsubmit version: {}".format(libsubmit.__version__))
# Update config with defaults
self._config = update_config(config, self.rundir)
# Set the data manager
if data_manager:
self.data_manager = data_manager
else:
self.data_manager = DataManager(config=self._config)
# Start the anonymized usage tracker and send init msg
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# Load checkpoints if any
cpts = self.load_checkpoints(checkpointFiles)
# Initialize the memoizer
self.memoizer = Memoizer(self, memoize=appCache, checkpoint=cpts)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
if self._config:
self._executors_managed = True
# Create the executors
epf = EPF()
self.executors = epf.make(self.rundir, self._config)
# set global vars from config
self.lazy_fail = self._config["globals"].get(
"lazyErrors", lazyErrors)
self.fail_retries = self._config["globals"].get("retries", retries)
self.flowcontrol = FlowControl(self, self._config)
self.checkpoint_mode = self._config["globals"].get(
"checkpointMode", checkpointMode)
if self.checkpoint_mode == "periodic":
period = self._config["globals"].get("checkpointPeriod",
"00:30:00")
try:
h, m, s = map(int, period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception as e:
logger.error(
"invalid checkpointPeriod provided:{0} expected HH:MM:SS"
.format(period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
else:
self._executors_managed = False
self.fail_retries = retries
self.lazy_fail = lazyErrors
self.executors = {i: x for i, x in enumerate(executors)}
self.flowcontrol = FlowNoControl(self, None)
self.checkpoint_mode = checkpointMode
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
self.task_launch_lock = threading.Lock()
logger.debug("Using executors: {0}".format(self.executors))
atexit.register(self.cleanup)
@staticmethod
def _count_deps(depends, task_id):
"""Internal.
Count the number of unresolved futures in the list depends.
"""
count = 0
for dep in depends:
if isinstance(dep, Future) or issubclass(type(dep), Future):
if not dep.done():
count += 1
return count
@property
def config(self):
"""Returns the fully initialized config that the DFK is actively using.
DO *NOT* update.
Returns:
- config (dict)
"""
return self._config
def handle_update(self, task_id, future, memo_cbk=False):
"""This function is called only as a callback from a task being done.
Move done task from runnable -> done
Move newly doable tasks from pending -> runnable , and launch
Args:
task_id (string) : Task id which is a uuid string
future (Future) : The future object corresponding to the task which
makes this callback
KWargs:
memo_cbk(Bool) : Indicates that the call is coming from a memo update,
that does not require additional memo updates.
"""
# TODO : Remove, this check is redundant
final_state_flag = False
if future.done():
try:
future.result()
except Exception as e:
logger.exception("Task {} failed".format(task_id))
# We keep the history separately, since the future itself could be
# tossed.
self.tasks[task_id]['fail_history'].append(future._exception)
self.tasks[task_id]['fail_count'] += 1
if not self.lazy_fail:
logger.debug("Eager fail, skipping retry logic")
raise e
if self.tasks[task_id]['fail_count'] <= self.fail_retries:
logger.debug("Task {} marked for retry".format(task_id))
self.tasks[task_id]['status'] = States.pending
else:
logger.info(
"Task {} failed after {} retry attempts".format(
task_id, self.fail_retries))
self.tasks[task_id]['status'] = States.failed
final_state_flag = True
else:
logger.info("Task {} completed".format(task_id))
self.tasks[task_id]['status'] = States.done
final_state_flag = True
if not memo_cbk and final_state_flag is True:
# Update the memoizer with the new result if this is not a
# result from a memo lookup and the task has reached a terminal state.
self.memoizer.update_memo(task_id, self.tasks[task_id], future)
if self.checkpoint_mode is 'task_exit':
self.checkpoint()
logger.debug(
"Task {} checkpoint created at task exit".format(task_id))
# Identify tasks that have resolved dependencies and launch
for tid in list(self.tasks):
# Skip all non-pending tasks
if self.tasks[tid]['status'] != States.pending:
continue
if self._count_deps(self.tasks[tid]['depends'], tid) == 0:
# We can now launch *task*
new_args, kwargs, exceptions = self.sanitize_and_wrap(
task_id, self.tasks[tid]['args'],
self.tasks[tid]['kwargs'])
self.tasks[tid]['args'] = new_args
self.tasks[tid]['kwargs'] = kwargs
if not exceptions:
# There are no dependency errors
exec_fu = None
# Acquire a lock, retest the state, launch
with self.task_launch_lock:
if self.tasks[tid]['status'] == States.pending:
self.tasks[tid]['status'] = States.running
exec_fu = self.launch_task(tid,
self.tasks[tid]['func'],
*new_args, **kwargs)
if exec_fu:
self.tasks[task_id]['exec_fu'] = exec_fu
try:
self.tasks[tid]['app_fu'].update_parent(exec_fu)
self.tasks[tid]['exec_fu'] = exec_fu
except AttributeError as e:
logger.error(
"Task {}: Caught AttributeError at update_parent"
.format(tid))
raise e
else:
logger.info(
"Task {} deferred due to dependency failure".format(
tid))
# Raise a dependency exception
self.tasks[tid]['status'] = States.dep_fail
try:
fu = Future()
fu.retries_left = 0
self.tasks[tid]['exec_fu'] = fu
self.tasks[tid]['app_fu'].update_parent(fu)
fu.set_exception(DependencyError(
exceptions, tid, None))
except AttributeError as e:
logger.error(
"Task {} AttributeError at update_parent".format(
tid))
raise e
return
def launch_task(self, task_id, executable, *args, **kwargs):
"""Handle the actual submission of the task to the executor layer.
If the app task has the sites attributes not set (default=='all')
the task is launched on a randomly selected executor from the
list of executors. This behavior could later be updates to support
binding to sites based on user specified criteria.
If the app task specifies a particular set of sites, it will be
targetted at those specific sites.
Args:
task_id (uuid string) : A uuid string that uniquely identifies the task
executable (callable) : A callable object
args (list of positional args)
kwargs (arbitrary keyword arguments)
Returns:
Future that tracks the execution of the submitted executable
"""
hit, memo_fu = self.memoizer.check_memo(task_id, self.tasks[task_id])
if hit:
self.handle_update(task_id, memo_fu, memo_cbk=True)
return memo_fu
target_sites = self.tasks[task_id]["sites"]
executor = None
if isinstance(target_sites, str) and target_sites.lower() == 'all':
# Pick a random site from the list
site, executor = random.choice(list(self.executors.items()))
elif isinstance(target_sites, list):
# Pick a random site from user specified list
try:
site = random.choice(target_sites)
executor = self.executors[site]
except Exception as e:
logger.error("Task {}: requests invalid site [{}]".format(
task_id, target_sites))
else:
logger.error(
"App {} specifies invalid site option, expects str|list".
format(self.tasks[task_id]['func'].__name__))
exec_fu = executor.submit(executable, *args, **kwargs)
exec_fu.retries_left = self.fail_retries - \
self.tasks[task_id]['fail_count']
exec_fu.add_done_callback(partial(self.handle_update, task_id))
logger.info("Task {} launched on site {}".format(task_id, site))
return exec_fu
@staticmethod
def _count_all_deps(task_id, args, kwargs):
"""Internal.
Count the number of unresolved futures in the list depends.
Args:
- task_id (uuid string) : Task_id
- args (List[args]) : The list of args list to the fn
- kwargs (Dict{kwargs}) : The dict of all kwargs passed to the fn
Returns:
- count, [list of dependencies]
"""
# Check the positional args
depends = []
count = 0
for dep in args:
if isinstance(dep, Future) or issubclass(dep.__class__, Future):
if not dep.done():
count += 1
depends.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future) or issubclass(dep.__class__, Future):
if not dep.done():
count += 1
depends.extend([dep])
# Check for futures in inputs=[<fut>...]
for dep in kwargs.get('inputs', []):
if issubclass(dep.__class__, Future) or isinstance(dep, Future):
if not dep.done():
count += 1
depends.extend([dep])
# logger.debug("Task:{0} dep_cnt:{1} deps:{2}".format(task_id, count, depends))
return count, depends
@staticmethod
def sanitize_and_wrap(task_id, args, kwargs):
"""This function should be called **ONLY** when all the futures we track have been resolved.
If the user hid futures a level below, we will not catch
it, and will (most likely) result in a type error .
Args:
task_id (uuid str) : Task id
func (Function) : App function
args (List) : Positional args to app function
kwargs (Dict) : Kwargs to app function
Return:
partial Function evaluated with all dependencies in args, kwargs and kwargs['inputs'] evaluated.
"""
dep_failures = []
# Replace item in args
new_args = []
for dep in args:
if isinstance(dep, Future) or issubclass(type(dep), Future):
try:
new_args.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_args.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future) or issubclass(type(dep), Future):
try:
kwargs[key] = dep.result()
except Exception as e:
dep_failures.extend([e])
# Check for futures in inputs=[<fut>...]
if 'inputs' in kwargs:
new_inputs = []
for dep in kwargs['inputs']:
if isinstance(dep, Future) or issubclass(type(dep), Future):
try:
new_inputs.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_inputs.extend([dep])
kwargs['inputs'] = new_inputs
return new_args, kwargs, dep_failures
def submit(self,
func,
*args,
parsl_sites='all',
fn_hash=None,
cache=False,
**kwargs):
"""Add task to the dataflow system.
>>> IF all deps are met :
>>> send to the runnable queue and launch the task
>>> ELSE:
>>> post the task in the pending queue
Args:
- func : A function object
- *args : Args to the function
KWargs :
- parsl_sites (List|String) : List of sites this call could go to.
Default='all'
- fn_hash (Str) : Hash of the function and inputs
Default=None
- cache (Bool) : To enable memoization or not
- kwargs (dict) : Rest of the kwargs to the fn passed as dict.
Returns:
(AppFuture) [DataFutures,]
"""
task_id = self.task_count
self.task_count += 1
# Get the dep count and a list of dependencies for the task
dep_cnt, depends = self._count_all_deps(task_id, args, kwargs)
task_def = {
'depends': depends,
'sites': parsl_sites,
'func': func,
'func_name': func.__name__,
'args': args,
'kwargs': kwargs,
'fn_hash': fn_hash,
'memoize': cache,
'callback': None,
'dep_cnt': dep_cnt,
'exec_fu': None,
'checkpoint': None,
'fail_count': 0,
'fail_history': [],
'env': None,
'status': States.unsched,
'app_fu': None
}
if task_id in self.tasks:
raise DuplicateTaskError(
"Task {0} in pending list".format(task_id))
else:
self.tasks[task_id] = task_def
# Extract stdout and stderr to pass to AppFuture:
task_stdout = kwargs.get('stdout', None)
task_stderr = kwargs.get('stderr', None)
logger.info("Task {} submitted for App {}, waiting on tasks {}".format(
task_id, task_def['func_name'], [fu.tid for fu in depends]))
# Handle three cases here:
# No pending deps
# - But has failures -> dep_fail
# - No failures -> running
# Has pending deps -> pending
if dep_cnt == 0:
new_args, kwargs, exceptions = self.sanitize_and_wrap(
task_id, args, kwargs)
self.tasks[task_id]['args'] = new_args
self.tasks[task_id]['kwargs'] = kwargs
if not exceptions:
self.tasks[task_id]['exec_fu'] = self.launch_task(
task_id, func, *new_args, **kwargs)
self.tasks[task_id]['app_fu'] = AppFuture(
self.tasks[task_id]['exec_fu'],
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['status'] = States.running
logger.debug("Task {} launched with AppFut:{}".format(
task_id, task_def['app_fu']))
else:
self.tasks[task_id]['exec_fu'] = None
app_fu = AppFuture(self.tasks[task_id]['exec_fu'],
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
app_fu.set_exception(
DependencyError(exceptions,
"Failures in input dependencies", None))
self.tasks[task_id]['app_fu'] = app_fu
self.tasks[task_id]['status'] = States.dep_fail
logger.debug(
"Task {} failed due to failure in parent task(s):{}".
format(task_id, task_def['app_fu']))
else:
# Send to pending, create the AppFuture with no parent and have it set
# when an executor future is available.
self.tasks[task_id]['app_fu'] = AppFuture(None,
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['status'] = States.pending
logger.debug("Task {} launched with AppFut:{}".format(
task_id, task_def['app_fu']))
return task_def['app_fu']
def cleanup(self):
"""DataFlowKernel cleanup.
This involves killing resources explicitly and sending die messages to IPP workers.
If the executors are managed, i.e created by the DFK
then : we scale_in each of the executors and call executor.shutdown
else : we do nothing. Executor cleanup is left to the user.
"""
logger.info("DFK cleanup initiated")
# Checkpointing takes priority over the rest of the tasks
# checkpoint if any valid checkpoint method is specified
if self.checkpoint_mode is not None:
self.checkpoint()
if self._checkpoint_timer:
logger.info("Stopping checkpoint timer")
self._checkpoint_timer.close()
# Send final stats
self.usage_tracker.send_message()
# We do not need to cleanup if the executors are managed outside
# the DFK
if not self._executors_managed:
return
logger.info("Terminating flow_control and strategy threads")
self.flowcontrol.close()
for executor in self.executors.values():
if executor.scaling_enabled:
job_ids = executor.execution_provider.resources.keys()
executor.scale_in(len(job_ids))
# We are not doing shutdown here because even with block=False this blocks.
executor.shutdown()
logger.info("DFK cleanup complete")
def checkpoint(self):
"""Checkpoint the dfk incrementally to a checkpoint file.
When called, every task that has been completed yet not
checkpointed is checkpointed to a file.
.. note::
Checkpointing only works if memoization is enabled
Returns:
Checkpoint dir if checkpoints were written successfully.
By default the checkpoints are written to the RUNDIR of the current
run under RUNDIR/checkpoints/{tasks.pkl, dfk.pkl}
"""
logger.info("Checkpointing.. ")
checkpoint_dir = '{0}/checkpoint'.format(self.rundir)
checkpoint_dfk = checkpoint_dir + '/dfk.pkl'
checkpoint_tasks = checkpoint_dir + '/tasks.pkl'
if not os.path.exists(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except FileExistsError as e:
pass
with open(checkpoint_dfk, 'wb') as f:
state = {
'config': self.config,
'rundir': self.rundir,
'task_count': self.task_count
}
pickle.dump(state, f)
count = 0
with open(checkpoint_tasks, 'wb+') as f:
for task_id in self.tasks:
if self.tasks[task_id]['app_fu'].done() and \
not self.tasks[task_id]['checkpoint']:
hashsum = self.tasks[task_id]['hashsum']
if not hashsum:
continue
t = {'hash': hashsum, 'exception': None, 'result': None}
try:
# Asking for the result will raise an exception if
# the app had failed. Should we even checkpoint these?
# TODO : Resolve this question ?
r = self.memoizer.hash_lookup(hashsum).result()
except Exception as e:
t['exception'] = e
else:
t['result'] = r
# We are using pickle here since pickle dumps to a file in 'w+'
# mode behave like a incremental log.
pickle.dump(t, f)
count += 1
self.tasks[task_id]['checkpoint'] = True
logger.debug("Task {} checkpointed".format(task_id))
self.checkpointed_tasks += count
if count == 0:
if self.checkpointed_tasks == 0:
logger.warn(
"No tasks checkpointed, please ensure caching is enabled")
else:
logger.debug("No tasks checkpointed")
else:
logger.info("Done checkpointing {} tasks".format(count))
return checkpoint_dir
def _load_checkpoints(self, checkpointDirs):
"""Load a checkpoint file into a lookup table.
The data being loaded from the pickle file mostly contains input
attributes of the task: func, args, kwargs, env...
To simplify the check of whether the exact task has been completed
in the checkpoint, we hash these input params and use it as the key
for the memoized lookup table.
Args:
- checkpointDirs (list) : List of filepaths to checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- memoized_lookup_table (dict)
"""
memo_lookup_table = {}
for checkpoint_dir in checkpointDirs:
checkpoint_file = os.path.join(checkpoint_dir, 'tasks.pkl')
try:
with open(checkpoint_file, 'rb') as f:
while True:
try:
data = pickle.load(f)
# Copy and hash only the input attributes
memo_fu = Future()
if data['exception']:
memo_fu.set_exception(data['exception'])
else:
memo_fu.set_result(data['result'])
memo_lookup_table[data['hash']] = memo_fu
except EOFError:
# Done with the checkpoint file
break
except FileNotFoundError:
reason = "Checkpoint file was not found: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
except Exception as e:
reason = "Failed to load Checkpoint: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
logger.info(
"Completed loading checkpoint:{0} with {1} tasks".format(
checkpoint_file, len(memo_lookup_table.keys())))
return memo_lookup_table
def load_checkpoints(self, checkpointDirs):
"""Load checkpoints from the checkpoint files into a dictionary.
The results are used to pre-populate the memoizer's lookup_table
Kwargs:
- checkpointDirs (list) : List of run folder to use as checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- dict containing, hashed -> future mappings
"""
self.memo_lookup_table = None
if not checkpointDirs:
return {}
if type(checkpointDirs) is not list:
raise BadCheckpoint("checkpointDirs expects a list of checkpoints")
return self._load_checkpoints(checkpointDirs)
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html'
)
self._config = config
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
self.run_dir = make_rundir(config.run_dir)
parsl.set_file_logger("{}/parsl.log".format(self.run_dir),
level=logging.DEBUG)
logger.info("Parsl version: {}".format(get_version()))
logger.info("Libsubmit version: {}".format(libsubmit.__version__))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self,
memoize=config.app_cache,
checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
data_manager = DataManager.get_data_manager(
max_threads=config.data_management_max_threads,
executors=config.executors)
self.executors = {
e.label: e
for e in config.executors + [data_manager]
}
for executor in self.executors.values():
executor.run_dir = self.run_dir # FIXME we should have a real interface for this
executor.start()
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception as e:
logger.error(
"invalid checkpoint_period provided:{0} expected HH:MM:SS".
format(period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
self.task_launch_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
class DataFlowKernel(object):
"""The DataFlowKernel adds dependency awareness to an existing executor.
It is responsible for managing futures, such that when dependencies are resolved,
pending tasks move to the runnable state.
Here is a simplified diagram of what happens internally::
User | DFK | Executor
----------------------------------------------------------
| |
Task-------+> +Submit |
App_Fu<------+--| |
| Dependencies met |
| task-------+--> +Submit
| Ex_Fu<------+----|
"""
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html'
)
self._config = config
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
self.run_dir = make_rundir(config.run_dir)
parsl.set_file_logger("{}/parsl.log".format(self.run_dir),
level=logging.DEBUG)
logger.info("Parsl version: {}".format(get_version()))
logger.info("Libsubmit version: {}".format(libsubmit.__version__))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self,
memoize=config.app_cache,
checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
data_manager = DataManager.get_data_manager(
max_threads=config.data_management_max_threads,
executors=config.executors)
self.executors = {
e.label: e
for e in config.executors + [data_manager]
}
for executor in self.executors.values():
executor.run_dir = self.run_dir # FIXME we should have a real interface for this
executor.start()
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception as e:
logger.error(
"invalid checkpoint_period provided:{0} expected HH:MM:SS".
format(period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
self.task_launch_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
@staticmethod
def _count_deps(depends, task_id):
"""Internal.
Count the number of unresolved futures in the list depends.
"""
count = 0
for dep in depends:
if isinstance(dep, Future) or issubclass(type(dep), Future):
if not dep.done():
count += 1
return count
@property
def config(self):
"""Returns the fully initialized config that the DFK is actively using.
DO *NOT* update.
Returns:
- config (dict)
"""
return self._config
def handle_update(self, task_id, future, memo_cbk=False):
"""This function is called only as a callback from a task being done.
Move done task from runnable -> done
Move newly doable tasks from pending -> runnable , and launch
Args:
task_id (string) : Task id which is a uuid string
future (Future) : The future object corresponding to the task which
makes this callback
KWargs:
memo_cbk(Bool) : Indicates that the call is coming from a memo update,
that does not require additional memo updates.
"""
final_state_flag = False
try:
res = future.result()
if isinstance(res, RemoteException):
res.reraise()
except Exception as e:
logger.exception("Task {} failed".format(task_id))
# We keep the history separately, since the future itself could be
# tossed.
self.tasks[task_id]['fail_history'].append(future._exception)
self.tasks[task_id]['fail_count'] += 1
if not self._config.lazy_errors:
logger.debug("Eager fail, skipping retry logic")
raise e
if self.tasks[task_id]['fail_count'] <= self._config.retries:
logger.debug("Task {} marked for retry".format(task_id))
self.tasks[task_id]['status'] = States.pending
else:
logger.info("Task {} failed after {} retry attempts".format(
task_id, self._config.retries))
self.tasks[task_id]['status'] = States.failed
final_state_flag = True
else:
logger.info("Task {} completed".format(task_id))
self.tasks[task_id]['status'] = States.done
final_state_flag = True
if not memo_cbk and final_state_flag is True:
# Update the memoizer with the new result if this is not a
# result from a memo lookup and the task has reached a terminal state.
self.memoizer.update_memo(task_id, self.tasks[task_id], future)
if self.checkpoint_mode is 'task_exit':
self.checkpoint(tasks=[task_id])
# Identify tasks that have resolved dependencies and launch
for tid in list(self.tasks):
# Skip all non-pending tasks
if self.tasks[tid]['status'] != States.pending:
continue
if self._count_deps(self.tasks[tid]['depends'], tid) == 0:
# We can now launch *task*
new_args, kwargs, exceptions = self.sanitize_and_wrap(
task_id, self.tasks[tid]['args'],
self.tasks[tid]['kwargs'])
self.tasks[tid]['args'] = new_args
self.tasks[tid]['kwargs'] = kwargs
if not exceptions:
# There are no dependency errors
exec_fu = None
# Acquire a lock, retest the state, launch
with self.task_launch_lock:
if self.tasks[tid]['status'] == States.pending:
self.tasks[tid]['status'] = States.running
exec_fu = self.launch_task(tid,
self.tasks[tid]['func'],
*new_args, **kwargs)
if exec_fu:
self.tasks[task_id]['exec_fu'] = exec_fu
try:
self.tasks[tid]['app_fu'].update_parent(exec_fu)
self.tasks[tid]['exec_fu'] = exec_fu
except AttributeError as e:
logger.error(
"Task {}: Caught AttributeError at update_parent"
.format(tid))
raise e
else:
logger.info(
"Task {} deferred due to dependency failure".format(
tid))
# Raise a dependency exception
self.tasks[tid]['status'] = States.dep_fail
try:
fu = Future()
fu.retries_left = 0
self.tasks[tid]['exec_fu'] = fu
self.tasks[tid]['app_fu'].update_parent(fu)
fu.set_exception(DependencyError(
exceptions, tid, None))
except AttributeError as e:
logger.error(
"Task {} AttributeError at update_parent".format(
tid))
raise e
return
def launch_task(self, task_id, executable, *args, **kwargs):
"""Handle the actual submission of the task to the executor layer.
If the app task has the executors attributes not set (default=='all')
the task is launched on a randomly selected executor from the
list of executors. This behavior could later be updated to support
binding to executors based on user specified criteria.
If the app task specifies a particular set of executors, it will be
targeted at those specific executors.
Args:
task_id (uuid string) : A uuid string that uniquely identifies the task
executable (callable) : A callable object
args (list of positional args)
kwargs (arbitrary keyword arguments)
Returns:
Future that tracks the execution of the submitted executable
"""
hit, memo_fu = self.memoizer.check_memo(task_id, self.tasks[task_id])
if hit:
self.handle_update(task_id, memo_fu, memo_cbk=True)
return memo_fu
executor_label = self.tasks[task_id]["executor"]
try:
executor = self.executors[executor_label]
except Exception as e:
logger.exception(
"Task {} requested invalid executor {}: config is\n{}".format(
task_id, executor_label, self._config))
exec_fu = executor.submit(executable, *args, **kwargs)
self.tasks[task_id]['status'] = States.running
exec_fu.retries_left = self._config.retries - \
self.tasks[task_id]['fail_count']
exec_fu.add_done_callback(partial(self.handle_update, task_id))
logger.info("Task {} launched on executor {}".format(
task_id, executor.label))
return exec_fu
def _add_input_deps(self, executor, args, kwargs):
"""Look for inputs of the app that are remote files. Submit stage_in
apps for such files and replace the file objects in the inputs list with
corresponding DataFuture objects.
Args:
- executor (str) : executor where the app is going to be launched
- args (List) : Positional args to app function
- kwargs (Dict) : Kwargs to app function
"""
inputs = kwargs.get('inputs', [])
for idx, f in enumerate(inputs):
if isinstance(f, File) and f.is_remote():
inputs[idx] = f.stage_in(executor)
@staticmethod
def _count_all_deps(task_id, args, kwargs):
"""Internal.
Count the number of unresolved futures in the list depends.
Args:
- task_id (uuid string) : Task_id
- args (List[args]) : The list of args list to the fn
- kwargs (Dict{kwargs}) : The dict of all kwargs passed to the fn
Returns:
- count, [list of dependencies]
"""
# Check the positional args
depends = []
count = 0
for dep in args:
if isinstance(dep, Future) or issubclass(dep.__class__, Future):
if not dep.done():
count += 1
depends.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future) or issubclass(dep.__class__, Future):
if not dep.done():
count += 1
depends.extend([dep])
# Check for futures in inputs=[<fut>...]
for dep in kwargs.get('inputs', []):
if issubclass(dep.__class__, Future) or isinstance(dep, Future):
if not dep.done():
count += 1
depends.extend([dep])
# logger.debug("Task:{0} dep_cnt:{1} deps:{2}".format(task_id, count, depends))
return count, depends
@staticmethod
def sanitize_and_wrap(task_id, args, kwargs):
"""This function should be called **ONLY** when all the futures we track have been resolved.
If the user hid futures a level below, we will not catch
it, and will (most likely) result in a type error .
Args:
task_id (uuid str) : Task id
func (Function) : App function
args (List) : Positional args to app function
kwargs (Dict) : Kwargs to app function
Return:
partial Function evaluated with all dependencies in args, kwargs and kwargs['inputs'] evaluated.
"""
dep_failures = []
# Replace item in args
new_args = []
for dep in args:
if isinstance(dep, Future) or issubclass(type(dep), Future):
try:
new_args.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_args.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future) or issubclass(type(dep), Future):
try:
kwargs[key] = dep.result()
except Exception as e:
dep_failures.extend([e])
# Check for futures in inputs=[<fut>...]
if 'inputs' in kwargs:
new_inputs = []
for dep in kwargs['inputs']:
if isinstance(dep, Future) or issubclass(type(dep), Future):
try:
new_inputs.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_inputs.extend([dep])
kwargs['inputs'] = new_inputs
return new_args, kwargs, dep_failures
def submit(self,
func,
*args,
executors='all',
fn_hash=None,
cache=False,
**kwargs):
"""Add task to the dataflow system.
If the app task has the executors attributes not set (default=='all')
the task will be launched on a randomly selected executor from the
list of executors. This behavior could later be updated to support
binding to executors based on user specified criteria.
If the app task specifies a particular set of executors, it will be
targetted at those specific executors.
>>> IF all deps are met:
>>> send to the runnable queue and launch the task
>>> ELSE:
>>> post the task in the pending queue
Args:
- func : A function object
- *args : Args to the function
KWargs :
- executors (list or string) : List of executors this call could go to.
Default='all'
- fn_hash (Str) : Hash of the function and inputs
Default=None
- cache (Bool) : To enable memoization or not
- kwargs (dict) : Rest of the kwargs to the fn passed as dict.
Returns:
(AppFuture) [DataFutures,]
"""
task_id = self.task_count
self.task_count += 1
if isinstance(executors, str) and executors.lower() == 'all':
choices = list(e for e in self.executors if e != 'data_manager')
elif isinstance(executors, list):
choices = executors
executor = random.choice(choices)
task_def = {
'depends': None,
'executor': executor,
'func': func,
'func_name': func.__name__,
'args': args,
'kwargs': kwargs,
'fn_hash': fn_hash,
'memoize': cache,
'callback': None,
'dep_cnt': None,
'exec_fu': None,
'checkpoint': None,
'fail_count': 0,
'fail_history': [],
'env': None,
'status': States.unsched,
'app_fu': None
}
if task_id in self.tasks:
raise DuplicateTaskError(
"Task {0} in pending list".format(task_id))
else:
self.tasks[task_id] = task_def
# Transform remote input files to data futures
self._add_input_deps(executor, args, kwargs)
# Get the dep count and a list of dependencies for the task
dep_cnt, depends = self._count_all_deps(task_id, args, kwargs)
self.tasks[task_id]['dep_cnt'] = dep_cnt
self.tasks[task_id]['depends'] = depends
# Extract stdout and stderr to pass to AppFuture:
task_stdout = kwargs.get('stdout')
task_stderr = kwargs.get('stderr')
logger.info("Task {} submitted for App {}, waiting on tasks {}".format(
task_id, task_def['func_name'], [fu.tid for fu in depends]))
# Handle three cases here:
# No pending deps
# - But has failures -> dep_fail
# - No failures -> running
# Has pending deps -> pending
if dep_cnt == 0:
new_args, kwargs, exceptions = self.sanitize_and_wrap(
task_id, args, kwargs)
self.tasks[task_id]['args'] = new_args
self.tasks[task_id]['kwargs'] = kwargs
if not exceptions:
self.tasks[task_id]['exec_fu'] = self.launch_task(
task_id, func, *new_args, **kwargs)
self.tasks[task_id]['app_fu'] = AppFuture(
self.tasks[task_id]['exec_fu'],
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
logger.debug("Task {} launched with AppFuture: {}".format(
task_id, task_def['app_fu']))
else:
fu = Future()
fu.set_exception(
DependencyError(exceptions,
"Failures in input dependencies", None))
fu.retries_left = 0
self.tasks[task_id]['exec_fu'] = fu
app_fu = AppFuture(self.tasks[task_id]['exec_fu'],
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['app_fu'] = app_fu
self.tasks[task_id]['status'] = States.dep_fail
logger.debug(
"Task {} failed due to failure in parent task(s):{}".
format(task_id, task_def['app_fu']))
else:
# Send to pending, create the AppFuture with no parent and have it set
# when an executor future is available.
self.tasks[task_id]['app_fu'] = AppFuture(None,
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['status'] = States.pending
logger.debug("Task {} launched with AppFuture: {}".format(
task_id, task_def['app_fu']))
return task_def['app_fu']
# it might also be interesting to assert that all DFK
# tasks are in a "final" state (3,4,5) when the DFK
# is closed down, and report some kind of warning.
# although really I'd like this to drain properly...
# and a drain function might look like this.
# If tasks have their states changed, this won't work properly
# but we can validate that...
def log_task_states(self):
logger.info("Summary of tasks in DFK:")
total_summarised = 0
keytasks = []
for tid in self.tasks:
keytasks.append((self.tasks[tid]['status'], tid))
def first(t):
return t[0]
sorted_keytasks = sorted(keytasks, key=first)
grouped_sorted_keytasks = itertools.groupby(sorted_keytasks, key=first)
# caution: g is an iterator that also advances the
# grouped_sorted_tasks iterator, so looping over
# both grouped_sorted_keytasks and g can only be done
# in certain patterns
for k, g in grouped_sorted_keytasks:
ts = []
for t in g:
tid = t[1]
ts.append(str(tid))
total_summarised = total_summarised + 1
tids_string = ", ".join(ts)
logger.info("Tasks in state {}: {}".format(str(k), tids_string))
total_in_tasks = len(self.tasks)
if total_summarised != total_in_tasks:
logger.error(
"Task count summarisation was inconsistent: summarised {} tasks, but tasks list contains {} tasks"
.format(total_summarised, total_in_tasks))
logger.info("End of summary")
def atexit_cleanup(self):
if not self.cleanup_called:
self.cleanup()
def cleanup(self):
"""DataFlowKernel cleanup.
This involves killing resources explicitly and sending die messages to IPP workers.
If the executors are managed (created by the DFK), then we call scale_in on each of
the executors and call executor.shutdown. Otherwise, we do nothing, and executor
cleanup is left to the user.
"""
logger.info("DFK cleanup initiated")
# this check won't detect two DFK cleanups happening from
# different threads extremely close in time because of
# non-atomic read/modify of self.cleanup_called
if self.cleanup_called:
raise Exception(
"attempt to clean up DFK when it has already been cleaned-up")
self.cleanup_called = True
self.log_task_states()
# Checkpointing takes priority over the rest of the tasks
# checkpoint if any valid checkpoint method is specified
if self.checkpoint_mode is not None:
self.checkpoint()
if self._checkpoint_timer:
logger.info("Stopping checkpoint timer")
self._checkpoint_timer.close()
# Send final stats
self.usage_tracker.send_message()
self.usage_tracker.close()
logger.info("Terminating flow_control and strategy threads")
self.flowcontrol.close()
for executor in self.executors.values():
if executor.managed:
if executor.scaling_enabled:
job_ids = executor.provider.resources.keys()
executor.scale_in(len(job_ids))
executor.shutdown()
logger.info("DFK cleanup complete")
def checkpoint(self, tasks=None):
"""Checkpoint the dfk incrementally to a checkpoint file.
When called, every task that has been completed yet not
checkpointed is checkpointed to a file.
Kwargs:
- tasks (List of task ids) : List of task ids to checkpoint. Default=None
if set to None, we iterate over all tasks held by the DFK.
.. note::
Checkpointing only works if memoization is enabled
Returns:
Checkpoint dir if checkpoints were written successfully.
By default the checkpoints are written to the RUNDIR of the current
run under RUNDIR/checkpoints/{tasks.pkl, dfk.pkl}
"""
with self.checkpoint_lock:
checkpoint_queue = None
if tasks:
checkpoint_queue = tasks
else:
checkpoint_queue = self.tasks
checkpoint_dir = '{0}/checkpoint'.format(self.run_dir)
checkpoint_dfk = checkpoint_dir + '/dfk.pkl'
checkpoint_tasks = checkpoint_dir + '/tasks.pkl'
if not os.path.exists(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except FileExistsError as e:
pass
with open(checkpoint_dfk, 'wb') as f:
state = {'rundir': self.run_dir, 'task_count': self.task_count}
pickle.dump(state, f)
count = 0
with open(checkpoint_tasks, 'ab') as f:
for task_id in checkpoint_queue:
if not self.tasks[task_id]['checkpoint'] and \
self.tasks[task_id]['status'] == States.done:
hashsum = self.tasks[task_id]['hashsum']
if not hashsum:
continue
t = {
'hash': hashsum,
'exception': None,
'result': None
}
try:
# Asking for the result will raise an exception if
# the app had failed. Should we even checkpoint these?
# TODO : Resolve this question ?
r = self.memoizer.hash_lookup(hashsum).result()
except Exception as e:
t['exception'] = e
else:
t['result'] = r
# We are using pickle here since pickle dumps to a file in 'ab'
# mode behave like a incremental log.
pickle.dump(t, f)
count += 1
self.tasks[task_id]['checkpoint'] = True
logger.debug("Task {} checkpointed".format(task_id))
self.checkpointed_tasks += count
if count == 0:
if self.checkpointed_tasks == 0:
logger.warn(
"No tasks checkpointed so far in this run. Please ensure caching is enabled"
)
else:
logger.debug("No tasks checkpointed in this pass.")
else:
logger.info("Done checkpointing {} tasks".format(count))
return checkpoint_dir
def _load_checkpoints(self, checkpointDirs):
"""Load a checkpoint file into a lookup table.
The data being loaded from the pickle file mostly contains input
attributes of the task: func, args, kwargs, env...
To simplify the check of whether the exact task has been completed
in the checkpoint, we hash these input params and use it as the key
for the memoized lookup table.
Args:
- checkpointDirs (list) : List of filepaths to checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- memoized_lookup_table (dict)
"""
memo_lookup_table = {}
for checkpoint_dir in checkpointDirs:
logger.info("Loading checkpoints from {}".format(checkpoint_dir))
checkpoint_file = os.path.join(checkpoint_dir, 'tasks.pkl')
try:
with open(checkpoint_file, 'rb') as f:
while True:
try:
data = pickle.load(f)
# Copy and hash only the input attributes
memo_fu = Future()
if data['exception']:
memo_fu.set_exception(data['exception'])
else:
memo_fu.set_result(data['result'])
memo_lookup_table[data['hash']] = memo_fu
except EOFError:
# Done with the checkpoint file
break
except FileNotFoundError:
reason = "Checkpoint file was not found: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
except Exception as e:
reason = "Failed to load checkpoint: {}".format(
checkpoint_file)
logger.error(reason)
raise BadCheckpoint(reason)
logger.info(
"Completed loading checkpoint:{0} with {1} tasks".format(
checkpoint_file, len(memo_lookup_table.keys())))
return memo_lookup_table
def load_checkpoints(self, checkpointDirs):
"""Load checkpoints from the checkpoint files into a dictionary.
The results are used to pre-populate the memoizer's lookup_table
Kwargs:
- checkpointDirs (list) : List of run folder to use as checkpoints
Eg. ['runinfo/001', 'runinfo/002']
Returns:
- dict containing, hashed -> future mappings
"""
self.memo_lookup_table = None
if not checkpointDirs:
return {}
if type(checkpointDirs) is not list:
raise BadCheckpoint("checkpointDirs expects a list of checkpoints")
return self._load_checkpoints(checkpointDirs)
def __init__(self,
config=None,
executors=None,
lazyErrors=True,
appCache=True,
rundir=None,
retries=0,
checkpointFiles=None):
""" Initialize the DataFlowKernel
Please note that keyword args passed to the DFK here will always override
options passed in via the config.
KWargs:
- config (Dict) : A single data object encapsulating all config attributes
- executors (list of Executor objs): Optional, kept for (somewhat) backward compatibility with 0.2.0
- lazyErrors(Bool) : Default=True, allow workflow to continue on app failures.
- appCache (Bool) :Enable caching of apps
- rundir (str) : Path to run directory. Defaults to ./runinfo/runNNN
- retries(int): Default=0, Set the number of retry attempts in case of failure
- checkpointFiles (list of str): List of filepaths to checkpoint files
Returns:
DataFlowKernel object
"""
# Create run dirs for this run
self.rundir = make_rundir(config=config, path=rundir)
parsl.set_file_logger("{}/parsl.log".format(self.rundir),
level=logging.INFO)
logger.info("Parsl version: {}".format(parsl.__version__))
logger.info("Libsubmit version: {}".format(libsubmit.__version__))
# Update config with defaults
self._config = update_config(config, self.rundir)
# Start the anonymized usage tracker and send init msg
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# Load checkpoints if any
cpts = self.load_checkpoints(checkpointFiles)
# Initialize the memoizer
self.memoizer = Memoizer(self, memoize=appCache, checkpoint=cpts)
if self._config:
self._executors_managed = True
# Create the executors
epf = EPF()
self.executors = epf.make(self.rundir, self._config)
# set global vars from config
self.lazy_fail = self._config["globals"].get(
"lazyErrors", lazyErrors)
self.fail_retries = self._config["globals"].get("retries", retries)
self.flowcontrol = FlowControl(self, self._config)
else:
self._executors_managed = False
self.fail_retries = retries
self.lazy_fail = lazyErrors
self.executors = {i: x for i, x in enumerate(executors)}
self.flowcontrol = FlowNoControl(self, None)
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
self.task_launch_lock = threading.Lock()
logger.debug("Using executors: {0}".format(self.executors))
atexit.register(self.cleanup)
def __init__(self, config=Config()):
"""Initialize the DataFlowKernel.
Parameters
----------
config : Config
A specification of all configuration options. For more details see the
:class:~`parsl.config.Config` documentation.
"""
# this will be used to check cleanup only happens once
self.cleanup_called = False
if isinstance(config, dict):
raise ConfigurationError(
'Expected `Config` class, received dictionary. For help, '
'see http://parsl.readthedocs.io/en/stable/stubs/parsl.config.Config.html'
)
self._config = config
logger.debug("Starting DataFlowKernel with config\n{}".format(config))
self.run_dir = make_rundir(config.run_dir)
parsl.set_file_logger("{}/parsl.log".format(self.run_dir),
level=logging.DEBUG)
logger.info("Parsl version: {}".format(get_version()))
logger.info("Libsubmit version: {}".format(libsubmit.__version__))
self.checkpoint_lock = threading.Lock()
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
# ES logging
self.db_logger_config = config.db_logger_config
self.db_logger = get_db_logger(
enable_es_logging=False
) if self.db_logger_config is None else get_db_logger(
**self.db_logger_config)
self.workflow_name = str(inspect.stack()[1][1])
self.time_began = datetime.now()
self.time_completed = None
self.run_id = self.workflow_name + "-" + str(self.time_began.minute)
self.dashboard = self.db_logger_config.get(
'dashboard_link',
None) if self.db_logger_config is not None else None
# TODO: make configurable
logger.info("Run id is: " + self.run_id)
if self.dashboard is not None:
logger.info("Dashboard is found at " + self.dashboard)
self.db_logger.info("Python version: {}".format(sys.version_info))
self.db_logger.info("Parsl version: {}".format(get_version()))
self.db_logger.info("Libsubmit version: {}".format(
libsubmit.__version__))
self.db_logger.info(
"DFK start",
extra={
"time_began":
str(self.time_began.strftime('%Y-%m-%d %H:%M:%S')),
'time_completed': str(self.time_completed),
'task_run_id': self.run_id,
'rundir': self.run_dir
})
self.db_logger.info("Name of script/workflow: " + self.run_id,
extra={'task_run_id': self.run_id})
for executor in self._config.executors:
self.db_logger.info("Listed executor: " + executor.label,
extra={'task_run_id': self.run_id})
# ES logging end
checkpoints = self.load_checkpoints(config.checkpoint_files)
self.memoizer = Memoizer(self,
memoize=config.app_cache,
checkpoint=checkpoints)
self.checkpointed_tasks = 0
self._checkpoint_timer = None
self.checkpoint_mode = config.checkpoint_mode
data_manager = DataManager.get_data_manager(
max_threads=config.data_management_max_threads,
executors=config.executors)
self.executors = {
e.label: e
for e in config.executors + [data_manager]
}
for executor in self.executors.values():
executor.run_dir = self.run_dir # FIXME we should have a real interface for this
executor.start()
if self.checkpoint_mode == "periodic":
try:
h, m, s = map(int, config.checkpoint_period.split(':'))
checkpoint_period = (h * 3600) + (m * 60) + s
self._checkpoint_timer = Timer(self.checkpoint,
interval=checkpoint_period)
except Exception as e:
logger.error(
"invalid checkpoint_period provided:{0} expected HH:MM:SS".
format(config.checkpoint_period))
self._checkpoint_timer = Timer(self.checkpoint,
interval=(30 * 60))
if any([x.managed for x in config.executors]):
self.flowcontrol = FlowControl(self)
else:
self.flowcontrol = FlowNoControl(self)
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
self.task_launch_lock = threading.Lock()
atexit.register(self.atexit_cleanup)
class DataFlowKernel(object):
""" DataFlowKernel
"""
def __init__(self, config=None, executors=None, lazy_fail=True,
rundir=None, fail_retries=2):
""" Initialize the DataFlowKernel
Please note that keyword args passed to the DFK here will always override
options passed in via the config.
KWargs:
config (Dict) : A single data object encapsulating all config attributes
executors (list of Executor objs): Optional, kept for (somewhat) backward compatibility with 0.2.0
lazy_fail(Bool) : Default=True, determine failure behavior
rundir (str) : Path to run directory. Defaults to ./runinfo/runNNN
fail_retries(int): Default=2, Set the number of retry attempts in case of failure
Returns:
DataFlowKernel object
"""
# Create run dirs for this run
self.rundir = make_rundir(config=config, path=rundir)
# Update config with defaults
self._config = update_config(config, self.rundir)
# Start the anonymized usage tracker and send init msg
self.usage_tracker = UsageTracker(self)
self.usage_tracker.send_message()
if self._config :
self._executors_managed = True
# Create the executors
epf = EPF()
self.executors = epf.make(self.rundir, self._config)
# set global vars from config
self.lazy_fail = self._config["globals"].get("lazyFail", lazy_fail)
self.fail_retries = self._config["globals"].get("fail_retries", fail_retries)
self.flowcontrol = FlowControl(self, self._config)
else:
self._executors_managed = False
self.fail_retries = fail_retries
self.lazy_fail = lazy_fail
self.executors = {i:x for i,x in enumerate(executors)}
print("Executors : ", self.executors)
self.flowcontrol = FlowNoControl(self, None)
self.task_count = 0
self.fut_task_lookup = {}
self.tasks = {}
logger.debug("Using executors: {0}".format(self.executors))
atexit.register(self.cleanup)
@staticmethod
def _count_deps(depends, task_id):
''' Internal. Count the number of unresolved futures in the list depends'''
count = 0
for dep in depends:
if isinstance(dep, Future) or issubclass(type(dep), Future):
logger.debug("Task[%s]: dep:%s done:%s", task_id, dep, dep.done())
if not dep.done():
count += 1
return count
@property
def config(self):
''' Returns the fully initialized config that the DFK is
actively using. DO *NOT* update.
Returns:
- config (dict)
'''
return self._config
def handle_update(self, task_id, future):
''' This function is called only as a callback from a task being done
Move done task from runnable -> done
Move newly doable tasks from pending -> runnable , and launch
Args:
task_id (string) : Task id which is a uuid string
future (Future) : The future object corresponding to the task which makes this callback
'''
if future.done():
# Untested
if not self.lazy_fail:
# Fail early
try:
future.result()
except Exception as e:
logger.warn("Exception : %s", future._exception)
logger.error("Task[%s]: FAILED with %s", task_id, future)
self.tasks[task_id]['status'] = States.failed
raise e
logger.debug("Task[%s]: COMPLETED with %s", task_id, future)
self.tasks[task_id]['status'] = States.done
# Identify tasks that have resolved dependencies and launch
for tid in list(self.tasks):
# Skip all non-pending tasks
if self.tasks[tid]['status'] != States.pending:
continue
if self._count_deps(self.tasks[tid]['depends'], tid) == 0:
# We can now launch *task*
new_args, kwargs, exceptions = self.sanitize_and_wrap(task_id,
self.tasks[tid]['args'],
self.tasks[tid]['kwargs'])
if not exceptions :
logger.debug("Task[%s] Launching Task".format(tid))
# There are no dependency errors
self.tasks[tid]['status'] = States.running
exec_fu = self.launch_task(tid, self.tasks[tid]['func'], *new_args, **kwargs)
self.tasks[task_id]['exec_fu'] = exec_fu
try:
self.tasks[tid]['app_fu'].update_parent(exec_fu)
self.tasks[tid]['exec_fu'] = exec_fu
except AttributeError as e:
logger.error("Task[%s]: Caught AttributeError at update_parent", tid)
raise e
else:
logger.debug("Task[%s]: Deferring Task due to dependency failure", tid)
# Raise a dependency exception
self.tasks[tid]['status'] = States.dep_fail
try:
fu = Future()
self.tasks[tid]['exec_fu'] = fu
self.tasks[tid]['app_fu'].update_parent(fu)
fu.set_exception(DependencyError(exceptions,
tid,
None))
print(self.tasks[tid]['app_fu'])
except AttributeError as e:
logger.error("Task[%s]: Caught AttributeError at update_parent", tid)
raise e
return
def write_status_log(self):
''' Write status log.
Args:
None
Kwargs:
None
'''
state_lens = {States.unsched : 0,
States.pending : 0,
States.runnable: 0,
States.running : 0,
States.done : 0,
States.failed : 0,
States.dep_fail: 0}
for tid in self.tasks:
state_lens[self.tasks[tid]['status']] += 1
logger.debug("Pending:%d Runnable:%d Done:%d", state_lens[States.pending],
state_lens[States.runnable],
state_lens[States.done])
def print_status_log(self):
''' Print status log in terms of pending, runnable and done tasks
Args:
None
Kwargs:
None
'''
state_lens = {States.unsched : 0,
States.pending : 0,
States.runnable: 0,
States.running : 0,
States.done : 0,
States.failed : 0,
States.dep_fail: 0}
for tid in self.tasks:
state_lens[self.tasks[tid]['status']] += 1
print("Pending:{0} Runnable:{1} Done:{2}".format( state_lens[States.pending],
state_lens[States.runnable],
state_lens[States.done] ))
def launch_task(self, task_id, executable, *args, **kwargs):
''' Handle the actual submission of the task to the executor layer
If the app task has the sites attributes not set (default=='all')
the task is launched on a randomly selected executor from the
list of executors. This behavior could later be updates to support
binding to sites based on user specified criteria.
If the app task specifies a particular set of sites, it will be
targetted at those specific sites.
Args:
task_id (uuid string) : A uuid string that uniquely identifies the task
executable (callable) : A callable object
args (list of positional args)
kwargs (list of keyword args)
Returns:
Future that tracks the execution of the submitted executable
'''
task_name = executable.__name__
target_sites = self.tasks[task_id]["sites"]
executor = None
if isinstance(target_sites, str) and target_sites.lower() == 'all' :
# Pick a random site from the list
site, executor = random.choice(list(self.executors.items()))
elif isinstance(target_sites, list) :
# Pick a random site from user specified list
try :
site = random.choice(target_sites)
executor = self.executors[site]
except Exception as e:
logger.error("Task[%s]: requests invalid site [%s]" % task_id, target_sites)
else:
logger.error("App[%s]: sites defined is invalid, neither str|list" % self.tasks[task_id]['func'].__name__)
exec_fu = executor.submit(executable, *args, **kwargs)
exec_fu.add_done_callback(partial(self.handle_update, task_id))
logger.debug("Task[%s] launched on executor:%s" %(task_id, executor))
return exec_fu
@staticmethod
def _count_all_deps(task_id, args, kwargs):
''' Internal. Count the number of unresolved futures in the list depends
Args:
task_id (uuid string) : Task_id
args (List[args]) : The list of args list to the fn
kwargs (Dict{kwargs}) : The dict of all kwargs passed to the fn
Returns:
count, [list of dependencies]
'''
# Check the positional args
depends = []
count = 0
for dep in args :
if isinstance(dep, Future) or issubclass(dep.__class__, Future):
if not dep.done():
count += 1
depends.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future) or issubclass(dep.__class__, Future):
if not dep.done():
count += 1
depends.extend([dep])
# Check for futures in inputs=[<fut>...]
for dep in kwargs.get('inputs', []):
if issubclass(dep.__class__, Future) or isinstance(dep, Future):
if not dep.done():
count += 1
depends.extend([dep])
#logger.debug("Task:{0} dep_cnt:{1} deps:{2}".format(task_id, count, depends))
return count, depends
@staticmethod
def sanitize_and_wrap(task_id, args, kwargs):
''' This function should be called **ONLY** when all the futures we track
have been resolved. If the user hid futures a level below, we will not catch
it, and will (most likely) result in a type error .
Args:
task_id (uuid str) : Task id
func (Function) : App function
args (List) : Positional args to app function
kwargs (Dict) : Kwargs to app function
Return:
partial Function evaluated with all dependencies in args, kwargs and kwargs['inputs'] evaluated.
'''
dep_failures = []
# Replace item in args
new_args = []
for dep in args :
if isinstance(dep, Future) or issubclass(type(dep), Future):
try :
new_args.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_args.extend([dep])
# Check for explicit kwargs ex, fu_1=<fut>
for key in kwargs:
dep = kwargs[key]
if isinstance(dep, Future) or issubclass(type(dep), Future):
try :
kwargs[key] = dep.result()
except Exception as e:
dep_failures.extend([e])
# Check for futures in inputs=[<fut>...]
if 'inputs' in kwargs:
new_inputs = []
for dep in kwargs['inputs']:
if isinstance(dep, Future) or issubclass(type(dep), Future):
try :
new_inputs.extend([dep.result()])
except Exception as e:
dep_failures.extend([e])
else:
new_inputs.extend([dep])
kwargs['inputs'] = new_inputs
return new_args, kwargs, dep_failures
def submit (self, func, *args, parsl_sites='all', **kwargs):
''' Add task to the dataflow system.
Args:
func : A function object
*args : Args to the function
KWargs :
Standard kwargs to the func as provided by the user
parsl_sites : List of sites as defined in the config, Default :'all'
This is the only kwarg that is passed in by the app definition.
If all deps are met :
send to the runnable queue
and launch the task
Else:
post the task in the pending queue
Returns:
(AppFuture) [DataFutures,]
'''
task_id = self.task_count
self.task_count += 1
# Get the dep count and a list of dependencies for the task
dep_cnt, depends = self._count_all_deps(task_id, args, kwargs)
task_def = { 'depends' : depends,
'sites' : parsl_sites,
'func' : func,
'func_name' : func.__name__,
'args' : args,
'kwargs' : kwargs,
'callback' : None,
'dep_cnt' : dep_cnt,
'exec_fu' : None,
'status' : States.unsched,
'app_fu' : None }
if task_id in self.tasks:
raise DuplicateTaskError("Task {0} in pending list".format(task_id))
else:
self.tasks[task_id] = task_def
# Extract stdout and stderr to pass to AppFuture:
task_stdout = kwargs.get('stdout', None)
task_stderr = kwargs.get('stderr', None)
if dep_cnt == 0 :
# Set to running
new_args, kwargs, exceptions = self.sanitize_and_wrap(task_id, args, kwargs)
if not exceptions:
self.tasks[task_id]['exec_fu'] = self.launch_task(task_id, func, *new_args, **kwargs)
self.tasks[task_id]['app_fu'] = AppFuture(self.tasks[task_id]['exec_fu'],
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['status'] = States.running
else:
self.tasks[task_id]['exec_fu'] = None
app_fu = AppFuture(self.tasks[task_id]['exec_fu'],
tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
app_fu.set_exception(DependencyError(exceptions,
"Failures in input dependencies",
None))
self.tasks[task_id]['app_fu'] = app_fu
self.tasks[task_id]['status'] = States.dep_fail
else:
# Send to pending, create the AppFuture with no parent and have it set
# when an executor future is available.
self.tasks[task_id]['app_fu'] = AppFuture(None, tid=task_id,
stdout=task_stdout,
stderr=task_stderr)
self.tasks[task_id]['status'] = States.pending
logger.debug("Task:%s Launched with AppFut:%s", task_id, task_def['app_fu'])
return task_def['app_fu']
def cleanup (self):
''' DataFlowKernel cleanup. This involves killing resources explicitly and
sending die messages to IPP workers.
If the executors are managed, i.e created by the DFK
then : we scale_in each of the executors and call executor.shutdown
else : we do nothing. Executor cleanup is left to the user.
'''
logger.debug("DFK cleanup initiated")
# Send final stats
self.usage_tracker.send_message()
# We do not need to cleanup if the executors are managed outside
# the DFK
if not self._executors_managed :
return
for executor in self.executors.values() :
if executor.scaling_enabled :
job_ids = executor.execution_provider.resources.keys()
executor.scale_in(len(job_ids))
# We are not doing shutdown here because even with block=False this blocks.
executor.shutdown()
logger.debug("DFK cleanup complete")