X-Git-Url: https://wannabe.guru.org/gitweb/?a=blobdiff_plain;f=executors.py;h=34528a33c2d10236cd7527fe53ffd027fa8020ca;hb=31c81f6539969a5eba864d3305f9fb7bf716a367;hp=46812c2b49203c2b23c021978e8e6fe334b80afa;hpb=36fea7f15ed17150691b5b3ead75450e575229ef;p=python_utils.git diff --git a/executors.py b/executors.py index 46812c2..34528a3 100644 --- a/executors.py +++ b/executors.py @@ -2,33 +2,32 @@ from __future__ import annotations -from abc import ABC, abstractmethod import concurrent.futures as fut -from collections import defaultdict -from dataclasses import dataclass import logging -import numpy import os import platform import random import subprocess import threading import time -from typing import Any, Callable, Dict, List, Optional, Set import warnings +from abc import ABC, abstractmethod +from collections import defaultdict +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Set import cloudpickle # type: ignore +import numpy from overrides import overrides -from ansi import bg, fg, underline, reset import argparse_utils import config -from decorator_utils import singleton -from exec_utils import run_silently, cmd_in_background, cmd_with_timeout import histogram as hist +from ansi import bg, fg, reset, underline +from decorator_utils import singleton +from exec_utils import cmd_in_background, cmd_with_timeout, run_silently from thread_utils import background_thread - logger = logging.getLogger(__name__) parser = config.add_commandline_args( @@ -74,22 +73,47 @@ def make_cloud_pickle(fun, *args, **kwargs): class BaseExecutor(ABC): def __init__(self, *, title=''): self.title = title - self.task_count = 0 self.histogram = hist.SimpleHistogram( hist.SimpleHistogram.n_evenly_spaced_buckets(int(0), int(500), 50) ) + self.task_count = 0 @abstractmethod def submit(self, function: Callable, *args, **kwargs) -> fut.Future: pass @abstractmethod - def shutdown(self, wait: bool = True) -> None: + def shutdown(self, *, wait: bool = True, quiet: bool = False) -> None: pass + def shutdown_if_idle(self, *, quiet: bool = False) -> bool: + """Shutdown the executor and return True if the executor is idle + (i.e. there are no pending or active tasks). Return False + otherwise. Note: this should only be called by the launcher + process. + + """ + if self.task_count == 0: + self.shutdown(wait=True, quiet=quiet) + return True + return False + def adjust_task_count(self, delta: int) -> None: + """Change the task count. Note: do not call this method from a + worker, it should only be called by the launcher process / + thread / machine. + + """ self.task_count += delta - logger.debug(f'Executor current task count is {self.task_count}') + logger.debug(f'Adjusted task count by {delta} to {self.task_count}') + + def get_task_count(self) -> int: + """Change the task count. Note: do not call this method from a + worker, it should only be called by the launcher process / + thread / machine. + + """ + return self.task_count class ThreadExecutor(BaseExecutor): @@ -104,34 +128,39 @@ class ThreadExecutor(BaseExecutor): self._thread_pool_executor = fut.ThreadPoolExecutor( max_workers=workers, thread_name_prefix="thread_executor_helper" ) + self.already_shutdown = False + # This is run on a different thread; do not adjust task count here. def run_local_bundle(self, fun, *args, **kwargs): logger.debug(f"Running local bundle at {fun.__name__}") - start = time.time() result = fun(*args, **kwargs) - end = time.time() - self.adjust_task_count(-1) - duration = end - start - logger.debug(f"{fun.__name__} finished; used {duration:.1f}s") - self.histogram.add_item(duration) return result @overrides def submit(self, function: Callable, *args, **kwargs) -> fut.Future: + if self.already_shutdown: + raise Exception('Submitted work after shutdown.') self.adjust_task_count(+1) newargs = [] newargs.append(function) for arg in args: newargs.append(arg) - return self._thread_pool_executor.submit( + start = time.time() + result = self._thread_pool_executor.submit( self.run_local_bundle, *newargs, **kwargs ) + result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start)) + result.add_done_callback(lambda _: self.adjust_task_count(-1)) + return result @overrides - def shutdown(self, wait=True) -> None: - logger.debug(f'Shutting down threadpool executor {self.title}') - print(self.histogram) - self._thread_pool_executor.shutdown(wait) + def shutdown(self, *, wait: bool = True, quiet: bool = False) -> None: + if not self.already_shutdown: + logger.debug(f'Shutting down threadpool executor {self.title}') + self._thread_pool_executor.shutdown(wait) + if not quiet: + print(self.histogram.__repr__(label_formatter='%ds')) + self.already_shutdown = True class ProcessExecutor(BaseExecutor): @@ -146,30 +175,35 @@ class ProcessExecutor(BaseExecutor): self._process_executor = fut.ProcessPoolExecutor( max_workers=workers, ) + self.already_shutdown = False + # This is run in another process; do not adjust task count here. def run_cloud_pickle(self, pickle): fun, args, kwargs = cloudpickle.loads(pickle) logger.debug(f"Running pickled bundle at {fun.__name__}") result = fun(*args, **kwargs) - self.adjust_task_count(-1) return result @overrides def submit(self, function: Callable, *args, **kwargs) -> fut.Future: + if self.already_shutdown: + raise Exception('Submitted work after shutdown.') start = time.time() self.adjust_task_count(+1) pickle = make_cloud_pickle(function, *args, **kwargs) result = self._process_executor.submit(self.run_cloud_pickle, pickle) - result.add_done_callback( - lambda _: self.histogram.add_item(time.time() - start) - ) + result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start)) + result.add_done_callback(lambda _: self.adjust_task_count(-1)) return result @overrides - def shutdown(self, wait=True) -> None: - logger.debug(f'Shutting down processpool executor {self.title}') - self._process_executor.shutdown(wait) - print(self.histogram) + def shutdown(self, *, wait: bool = True, quiet: bool = False) -> None: + if not self.already_shutdown: + logger.debug(f'Shutting down processpool executor {self.title}') + self._process_executor.shutdown(wait) + if not quiet: + print(self.histogram.__repr__(label_formatter='%ds')) + self.already_shutdown = True def __getstate__(self): state = self.__dict__.copy() @@ -213,7 +247,7 @@ class BundleDetails: end_ts: float slower_than_local_p95: bool slower_than_global_p95: bool - src_bundle: BundleDetails + src_bundle: Optional[BundleDetails] is_cancelled: threading.Event was_cancelled: bool backup_bundles: Optional[List[BundleDetails]] @@ -253,14 +287,12 @@ class RemoteExecutorStatus: self.worker_count: int = total_worker_count self.known_workers: Set[RemoteWorkerRecord] = set() self.start_time: float = time.time() - self.start_per_bundle: Dict[str, float] = defaultdict(float) + self.start_per_bundle: Dict[str, Optional[float]] = defaultdict(float) self.end_per_bundle: Dict[str, float] = defaultdict(float) self.finished_bundle_timings_per_worker: Dict[ RemoteWorkerRecord, List[float] ] = {} - self.in_flight_bundles_by_worker: Dict[ - RemoteWorkerRecord, Set[str] - ] = {} + self.in_flight_bundles_by_worker: Dict[RemoteWorkerRecord, Set[str]] = {} self.bundle_details_by_uuid: Dict[str, BundleDetails] = {} self.finished_bundle_timings: List[float] = [] self.last_periodic_dump: Optional[float] = None @@ -270,9 +302,7 @@ class RemoteExecutorStatus: # as a memory fence for modifications to bundle. self.lock: threading.Lock = threading.Lock() - def record_acquire_worker( - self, worker: RemoteWorkerRecord, uuid: str - ) -> None: + def record_acquire_worker(self, worker: RemoteWorkerRecord, uuid: str) -> None: with self.lock: self.record_acquire_worker_already_locked(worker, uuid) @@ -290,9 +320,7 @@ class RemoteExecutorStatus: with self.lock: self.record_bundle_details_already_locked(details) - def record_bundle_details_already_locked( - self, details: BundleDetails - ) -> None: + def record_bundle_details_already_locked(self, details: BundleDetails) -> None: assert self.lock.locked() self.bundle_details_by_uuid[details.uuid] = details @@ -303,9 +331,7 @@ class RemoteExecutorStatus: was_cancelled: bool, ) -> None: with self.lock: - self.record_release_worker_already_locked( - worker, uuid, was_cancelled - ) + self.record_release_worker_already_locked(worker, uuid, was_cancelled) def record_release_worker_already_locked( self, @@ -318,7 +344,9 @@ class RemoteExecutorStatus: self.end_per_bundle[uuid] = ts self.in_flight_bundles_by_worker[worker].remove(uuid) if not was_cancelled: - bundle_latency = ts - self.start_per_bundle[uuid] + start = self.start_per_bundle[uuid] + assert start + bundle_latency = ts - start x = self.finished_bundle_timings_per_worker.get(worker, list()) x.append(bundle_latency) self.finished_bundle_timings_per_worker[worker] = x @@ -377,11 +405,7 @@ class RemoteExecutorStatus: ret += f' ...{in_flight} bundles currently in flight:\n' for bundle_uuid in self.in_flight_bundles_by_worker[worker]: details = self.bundle_details_by_uuid.get(bundle_uuid, None) - pid = ( - str(details.pid) - if (details and details.pid != 0) - else "TBD" - ) + pid = str(details.pid) if (details and details.pid != 0) else "TBD" if self.start_per_bundle[bundle_uuid] is not None: sec = ts - self.start_per_bundle[bundle_uuid] ret += f' (pid={pid}): {details} for {sec:.1f}s so far ' @@ -412,10 +436,7 @@ class RemoteExecutorStatus: assert self.lock.locked() self.total_bundles_submitted = total_bundles_submitted ts = time.time() - if ( - self.last_periodic_dump is None - or ts - self.last_periodic_dump > 5.0 - ): + if self.last_periodic_dump is None or ts - self.last_periodic_dump > 5.0: print(self) self.last_periodic_dump = ts @@ -429,9 +450,7 @@ class RemoteWorkerSelectionPolicy(ABC): pass @abstractmethod - def acquire_worker( - self, machine_to_avoid=None - ) -> Optional[RemoteWorkerRecord]: + def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]: pass @@ -444,26 +463,32 @@ class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): return False @overrides - def acquire_worker( - self, machine_to_avoid=None - ) -> Optional[RemoteWorkerRecord]: + def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]: grabbag = [] for worker in self.workers: - for x in range(0, worker.count): - for y in range(0, worker.weight): - grabbag.append(worker) - - for _ in range(0, 5): - random.shuffle(grabbag) - worker = grabbag[0] - if worker.machine != machine_to_avoid or _ > 2: + if worker.machine != machine_to_avoid: if worker.count > 0: - worker.count -= 1 - logger.debug(f'Selected worker {worker}') - return worker - msg = 'Unexpectedly could not find a worker, retrying...' - logger.warning(msg) - return None + for _ in range(worker.count * worker.weight): + grabbag.append(worker) + + if len(grabbag) == 0: + logger.debug( + f'There are no available workers that avoid {machine_to_avoid}...' + ) + for worker in self.workers: + if worker.count > 0: + for _ in range(worker.count * worker.weight): + grabbag.append(worker) + + if len(grabbag) == 0: + logger.warning('There are no available workers?!') + return None + + worker = random.sample(grabbag, 1)[0] + assert worker.count > 0 + worker.count -= 1 + logger.debug(f'Chose worker {worker}') + return worker class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): @@ -534,9 +559,10 @@ class RemoteExecutor(BaseExecutor): self.heartbeat_thread, self.heartbeat_stop_event, ) = self.run_periodic_heartbeat() + self.already_shutdown = False @background_thread - def run_periodic_heartbeat(self, stop_event) -> None: + def run_periodic_heartbeat(self, stop_event: threading.Event) -> None: while not stop_event.is_set(): time.sleep(5.0) logger.debug('Running periodic heartbeat code...') @@ -544,8 +570,10 @@ class RemoteExecutor(BaseExecutor): logger.debug('Periodic heartbeat thread shutting down.') def heartbeat(self) -> None: + # Note: this is invoked on a background thread, not an + # executor thread. Be careful what you do with it b/c it + # needs to get back and dump status again periodically. with self.status.lock: - # Dump regular progress report self.status.periodic_dump(self.total_bundles_submitted) # Look for bundles to reschedule via executor.submit @@ -560,7 +588,7 @@ class RemoteExecutor(BaseExecutor): if ( num_done > 2 and num_idle_workers > 1 - and (self.last_backup is None or (now - self.last_backup > 6.0)) + and (self.last_backup is None or (now - self.last_backup > 9.0)) and self.backup_lock.acquire(blocking=False) ): try: @@ -585,9 +613,7 @@ class RemoteExecutor(BaseExecutor): break for uuid in bundle_uuids: - bundle = self.status.bundle_details_by_uuid.get( - uuid, None - ) + bundle = self.status.bundle_details_by_uuid.get(uuid, None) if ( bundle is not None and bundle.src_bundle is None @@ -678,9 +704,7 @@ class RemoteExecutor(BaseExecutor): logger.critical(msg) raise Exception(msg) - def release_worker( - self, bundle: BundleDetails, *, was_cancelled=True - ) -> None: + def release_worker(self, bundle: BundleDetails, *, was_cancelled=True) -> None: worker = bundle.worker assert worker is not None logger.debug(f'Released worker {worker}') @@ -764,14 +788,14 @@ class RemoteExecutor(BaseExecutor): # Send input code / data to worker machine if it's not local. if hostname not in machine: try: - cmd = f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}' + cmd = ( + f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}' + ) start_ts = time.time() logger.info(f"{bundle}: Copying work to {worker} via {cmd}.") run_silently(cmd) xfer_latency = time.time() - start_ts - logger.debug( - f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s." - ) + logger.debug(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.") except Exception as e: self.release_worker(bundle) if is_original: @@ -804,9 +828,7 @@ class RemoteExecutor(BaseExecutor): f' /home/scott/lib/python_modules/remote_worker.py' f' --code_file {bundle.code_file} --result_file {bundle.result_file}"' ) - logger.debug( - f'{bundle}: Executing {cmd} in the background to kick off work...' - ) + logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...') p = cmd_in_background(cmd, silent=True) bundle.pid = p.pid logger.debug( @@ -815,9 +837,10 @@ class RemoteExecutor(BaseExecutor): return self.wait_for_process(p, bundle, 0) def wait_for_process( - self, p: subprocess.Popen, bundle: BundleDetails, depth: int + self, p: Optional[subprocess.Popen], bundle: BundleDetails, depth: int ) -> Any: machine = bundle.machine + assert p pid = p.pid if depth > 3: logger.error( @@ -935,9 +958,7 @@ class RemoteExecutor(BaseExecutor): # Re-raise the exception; the code in wait_for_process may # decide to emergency_retry_nasty_bundle here. raise Exception(e) - logger.debug( - f'Removing local (master) {code_file} and {result_file}.' - ) + logger.debug(f'Removing local (master) {code_file} and {result_file}.') os.remove(f'{result_file}') os.remove(f'{code_file}') @@ -962,10 +983,12 @@ class RemoteExecutor(BaseExecutor): # Tell the original to stop if we finished first. if not was_cancelled: + orig_bundle = bundle.src_bundle + assert orig_bundle logger.debug( - f'{bundle}: Notifying original {bundle.src_bundle.uuid} we beat them to it.' + f'{bundle}: Notifying original {orig_bundle.uuid} we beat them to it.' ) - bundle.src_bundle.is_cancelled.set() + orig_bundle.is_cancelled.set() self.release_worker(bundle, was_cancelled=was_cancelled) return result @@ -1049,7 +1072,9 @@ class RemoteExecutor(BaseExecutor): # they will move the result_file to this machine and let # the original pick them up and unpickle them. - def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future: + def emergency_retry_nasty_bundle( + self, bundle: BundleDetails + ) -> Optional[fut.Future]: is_original = bundle.src_bundle is None bundle.worker = None avoid_last_machine = bundle.machine @@ -1082,18 +1107,23 @@ class RemoteExecutor(BaseExecutor): @overrides def submit(self, function: Callable, *args, **kwargs) -> fut.Future: + if self.already_shutdown: + raise Exception('Submitted work after shutdown.') pickle = make_cloud_pickle(function, *args, **kwargs) bundle = self.create_original_bundle(pickle, function.__name__) self.total_bundles_submitted += 1 return self._helper_executor.submit(self.launch, bundle) @overrides - def shutdown(self, wait=True) -> None: - logging.debug(f'Shutting down RemoteExecutor {self.title}') - self.heartbeat_stop_event.set() - self.heartbeat_thread.join() - self._helper_executor.shutdown(wait) - print(self.histogram) + def shutdown(self, *, wait: bool = True, quiet: bool = False) -> None: + if not self.already_shutdown: + logging.debug(f'Shutting down RemoteExecutor {self.title}') + self.heartbeat_stop_event.set() + self.heartbeat_thread.join() + self._helper_executor.shutdown(wait) + if not quiet: + print(self.histogram.__repr__(label_formatter='%ds')) + self.already_shutdown = True @singleton @@ -1133,7 +1163,7 @@ class DefaultExecutors(object): RemoteWorkerRecord( username='scott', machine='cheetah.house', - weight=34, + weight=30, count=6, ), ) @@ -1163,7 +1193,7 @@ class DefaultExecutors(object): RemoteWorkerRecord( username='scott', machine='puma.cabin', - weight=25, + weight=30, count=6, ), ) @@ -1173,7 +1203,7 @@ class DefaultExecutors(object): RemoteWorkerRecord( username='scott', machine='backup.house', - weight=7, + weight=8, count=2, ), ) @@ -1191,11 +1221,11 @@ class DefaultExecutors(object): def shutdown(self) -> None: if self.thread_executor is not None: - self.thread_executor.shutdown() + self.thread_executor.shutdown(wait=True, quiet=True) self.thread_executor = None if self.process_executor is not None: - self.process_executor.shutdown() + self.process_executor.shutdown(wait=True, quiet=True) self.process_executor = None if self.remote_executor is not None: - self.remote_executor.shutdown() + self.remote_executor.shutdown(wait=True, quiet=True) self.remote_executor = None