1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
|
from __future__ import annotations
import datetime
import logging
import threading
import time
from abc import abstractmethod, abstractproperty
from concurrent.futures import Future
from itertools import compress
from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
from parsl.executors.base import ParslExecutor
from parsl.executors.errors import BadStateException, ScalingFailed
from parsl.jobs.error_handlers import noop_error_handler, simple_error_handler
from parsl.jobs.states import TERMINAL_STATES, JobState, JobStatus
from parsl.monitoring.message_type import MessageType
from parsl.providers.base import ExecutionProvider
from parsl.utils import AtomicIDCounter
logger = logging.getLogger(__name__)
class BlockProviderExecutor(ParslExecutor):
"""A base class for executors which scale using blocks.
This base class is intended to help with executors which:
- use blocks of workers to execute tasks
- blocks of workers are launched on a batch system through
an `ExecutionProvider`
An implementing class should implement the abstract methods required by
`ParslExecutor` to submit tasks, as well as BlockProviderExecutor
abstract methods to provide the executor-specific command to start a block
of workers (the ``_get_launch_command`` method), and some basic scaling
information (``outstanding`` and ``workers_per_node`` properties).
This base class provides a ``scale_out`` method which will launch new
blocks. It does not provide a ``scale_in`` method, because scale-in
behaviour is not well defined in the Parsl scaling model and so behaviour
is left to individual executors.
Parsl scaling will provide scaling between min_blocks and max_blocks by
invoking scale_out, but it will not initialize the blocks requested by
any init_blocks parameter. Subclasses must implement that behaviour
themselves.
"""
def __init__(self, *,
provider: Optional[ExecutionProvider],
block_error_handler: Union[bool, Callable[[BlockProviderExecutor, Dict[str, JobStatus]], None]]):
super().__init__()
self._provider = provider
self.block_error_handler: Callable[[BlockProviderExecutor, Dict[str, JobStatus]], None]
if isinstance(block_error_handler, bool):
if block_error_handler:
self.block_error_handler = simple_error_handler
else:
self.block_error_handler = noop_error_handler
else:
self.block_error_handler = block_error_handler
self._executor_bad_state = threading.Event()
self._executor_exception: Optional[Exception] = None
self._block_id_counter = AtomicIDCounter()
self._tasks = {} # type: Dict[object, Future]
self._last_poll_time = 0.0
# these four structures track, in loosely coordinated fashion, the
# existence of blocks and jobs and how to map between their
# identifiers.
self.blocks_to_job_id = {} # type: Dict[str, str]
self.job_ids_to_block = {} # type: Dict[str, str]
# errors can happen during the submit call to the provider; this is used
# to keep track of such errors so that they can be handled in one place
# together with errors reported by status()
self._simulated_status: Dict[str, JobStatus] = {}
# this stores an approximation (sometimes delayed) of the latest status
# of pending, active and recently terminated blocks
self._status = {} # type: Dict[str, JobStatus]
def _make_status_dict(self, block_ids: List[str], status_list: List[JobStatus]) -> Dict[str, JobStatus]:
"""Given a list of block ids and a list of corresponding status strings,
returns a dictionary mapping each block id to the corresponding status
:param block_ids: the list of block ids
:param status_list: the list of job status strings
:return: the resulting dictionary
"""
if len(block_ids) != len(status_list):
raise IndexError("block id list and status string list differ in size")
d = {}
for i in range(len(block_ids)):
d[block_ids[i]] = status_list[i]
return d
@property
def status_polling_interval(self):
"""Returns the interval, in seconds, at which the status method should be called. The
assumption here is that, once initialized, an executor's polling interval is fixed.
In practice, at least given the current situation, the executor uses a single task provider
and this method is a delegate to the corresponding method in the provider.
:return: the number of seconds to wait between calls to status() or zero if no polling
should be done
"""
if self._provider is None:
return 0
else:
return self._provider.status_polling_interval
@abstractproperty
def outstanding(self) -> int:
"""This should return the number of tasks that the executor has been given to run (waiting to run, and running now)"""
raise NotImplementedError("Classes inheriting from BlockProviderExecutor must implement "
"outstanding()")
def set_bad_state_and_fail_all(self, exception: Exception):
"""Allows external error handlers to mark this executor as irrecoverably bad and cause
all tasks submitted to it now and in the future to fail. The executor is responsible
for checking :method:bad_state_is_set() in the :method:submit() method and raising the
appropriate exception, which is available through :method:executor_exception().
"""
logger.exception("Setting bad state due to exception", exc_info=exception)
self._executor_exception = exception
# Set bad state to prevent new tasks from being submitted
self._executor_bad_state.set()
# We set all current tasks to this exception to make sure that
# this is raised in the main context.
for task in self._tasks:
self._tasks[task].set_exception(BadStateException(self, self._executor_exception))
@property
def bad_state_is_set(self):
"""Returns true if this executor is in an irrecoverable error state. If this method
returns true, :property:executor_exception should contain an exception indicating the
cause.
"""
return self._executor_bad_state.is_set()
@property
def executor_exception(self):
"""Returns an exception that indicates why this executor is in an irrecoverable state."""
return self._executor_exception
def handle_errors(self, status: Dict[str, JobStatus]) -> None:
"""This method is called by the error management infrastructure after a status poll. The
executor implementing this method is then responsible for detecting abnormal conditions
based on the status of submitted jobs. If the executor does not implement any special
error handling, this method should return False, in which case a generic error handling
scheme will be used.
:param status: status of all jobs launched by this executor
"""
self.block_error_handler(self, status)
@property
def tasks(self) -> Dict[object, Future]:
return self._tasks
@property
def provider(self):
return self._provider
def _filter_scale_in_ids(self, to_kill: Sequence[Any], killed: Sequence[bool]) -> Sequence[Any]:
""" Filter out job id's that were not killed
"""
assert len(to_kill) == len(killed)
if False in killed:
killed_job_ids = [jid for jid, k in zip(to_kill, killed) if k]
not_killed_job_ids = [jid for jid, k in zip(to_kill, killed) if not k]
logger.warning("Some jobs were not killed successfully: "
f"killed jobs: {killed_job_ids}, "
f"not-killed jobs: {not_killed_job_ids}")
# Filters first iterable by bool values in second
return list(compress(to_kill, killed))
def scale_out_facade(self, n: int) -> List[str]:
"""Scales out the number of blocks by "blocks"
"""
if not self.provider:
raise ScalingFailed(self, "No execution provider available")
block_ids = []
monitoring_status_changes = {}
logger.info(f"Scaling out by {n} blocks")
for _ in range(n):
block_id = str(self._block_id_counter.get_id())
logger.info(f"Allocated block ID {block_id}")
try:
job_id = self._launch_block(block_id)
pending_status = JobStatus(JobState.PENDING)
self.blocks_to_job_id[block_id] = job_id
self.job_ids_to_block[job_id] = block_id
self._status[block_id] = pending_status
monitoring_status_changes[block_id] = pending_status
block_ids.append(block_id)
except Exception as ex:
failed_status = JobStatus(JobState.FAILED, "Failed to start block {}: {}".format(block_id, ex))
self._simulated_status[block_id] = failed_status
self._status[block_id] = failed_status
self.send_monitoring_info(monitoring_status_changes)
return block_ids
def scale_in(self, blocks: int) -> List[str]:
"""Scale in method.
Cause the executor to reduce the number of blocks by count.
The default implementation will kill blocks without regard to their
status or whether they are executing tasks. Executors with more
nuanced scaling strategies might overload this method to work with
that strategy - see the HighThroughputExecutor for an example of that.
:return: A list of block ids corresponding to the blocks that were removed.
"""
active_blocks = [block_id for block_id, status in self._status.items()
if status.state not in TERMINAL_STATES]
block_ids_to_kill = active_blocks[:blocks]
job_ids_to_kill = [self.blocks_to_job_id[block] for block in block_ids_to_kill]
# Cancel the blocks provisioned
if self.provider:
logger.info(f"Scaling in jobs: {job_ids_to_kill}")
r = self.provider.cancel(job_ids_to_kill)
job_ids = self._filter_scale_in_ids(job_ids_to_kill, r)
block_ids_killed = [self.job_ids_to_block[job_id] for job_id in job_ids]
return block_ids_killed
else:
logger.error("No execution provider available to scale in")
return []
def _launch_block(self, block_id: str) -> Any:
launch_cmd = self._get_launch_command(block_id)
job_name = f"parsl.{self.label}.block-{block_id}"
logger.debug("Submitting to provider with job_name %s", job_name)
job_id = self.provider.submit(launch_cmd, 1, job_name)
if job_id:
logger.debug(f"Launched block {block_id} on executor {self.label} with job ID {job_id}")
else:
raise ScalingFailed(self,
"Attempt to provision nodes did not return a job ID")
return job_id
@abstractmethod
def _get_launch_command(self, block_id: str) -> str:
pass
def _get_block_and_job_ids(self) -> Tuple[List[str], List[Any]]:
# Not using self.blocks.keys() and self.blocks.values() simultaneously
# The dictionary may be changed during invoking this function
# As scale_in and scale_out are invoked in multiple threads
block_ids = list(self.blocks_to_job_id.keys())
job_ids = [] # types: List[Any]
for bid in block_ids:
job_ids.append(self.blocks_to_job_id[bid])
return block_ids, job_ids
@abstractproperty
def workers_per_node(self) -> Union[int, float]:
pass
def send_monitoring_info(self, status: Dict) -> None:
# Send monitoring info for HTEX when monitoring enabled
if self.submit_monitoring_radio:
msg = self.create_monitoring_info(status)
logger.debug("Sending block monitoring message: %r", msg)
self.submit_monitoring_radio.send((MessageType.BLOCK_INFO, msg))
def create_monitoring_info(self, status: Dict[str, JobStatus]) -> Sequence[object]:
"""Create a monitoring message for each block based on the poll status.
"""
msg = []
for bid, s in status.items():
d: Dict[str, Any] = {}
d['run_id'] = self.run_id
d['status'] = s.status_name
d['timestamp'] = datetime.datetime.now()
d['executor_label'] = self.label
d['job_id'] = self.blocks_to_job_id.get(bid, None)
d['block_id'] = bid
msg.append(d)
return msg
def poll_facade(self) -> None:
now = time.time()
if now >= self._last_poll_time + self.status_polling_interval:
previous_status = self._status
self._status = self.status()
self._last_poll_time = now
delta_status = {}
for block_id in self._status:
if block_id not in previous_status \
or previous_status[block_id].state != self._status[block_id].state:
delta_status[block_id] = self._status[block_id]
if delta_status:
self.send_monitoring_info(delta_status)
def status(self) -> Dict[str, JobStatus]:
"""Return the status of all jobs/blocks currently known to this executor.
:return: a dictionary mapping block ids (in string) to job status
"""
if self._provider:
block_ids, job_ids = self._get_block_and_job_ids()
status = self._make_status_dict(block_ids, self._provider.status(job_ids))
else:
status = {}
status.update(self._simulated_status)
return status
@property
def status_facade(self) -> Dict[str, JobStatus]:
"""Return the status of all jobs/blocks of the executor of this poller.
:return: a dictionary mapping block ids (in string) to job status
"""
return self._status
def scale_in_facade(self, n: int, max_idletime: Optional[float] = None) -> List[str]:
if max_idletime is None:
block_ids = self.scale_in(n)
else:
# This is a HighThroughputExecutor-specific interface violation.
# This code hopes, through pan-codebase reasoning, that this
# scale_in method really does come from HighThroughputExecutor,
# and so does have an extra max_idletime parameter not present
# in the executor interface.
block_ids = self.scale_in(n, max_idletime=max_idletime) # type: ignore[call-arg]
if block_ids is not None:
new_status = {}
for block_id in block_ids:
logger.debug("Marking block %s as SCALED_IN", block_id)
s = JobStatus(JobState.SCALED_IN)
new_status[block_id] = s
self._status[block_id] = s
self._simulated_status[block_id] = s
self.send_monitoring_info(new_status)
return block_ids
|