X-Git-Url: https://wannabe.guru.org/gitweb/?a=blobdiff_plain;ds=inline;f=executors.py;h=6723bb9e2df4f1990f9780f15a813e6a4b3f0548;hb=957c962294cde93c4c649ed18dedd16df071878d;hp=63efd812e54b5b796daff892f9ff1e75fd4e6569;hpb=b10d30a46e601c9ee1f843241f2d69a1f90f7a94;p=python_utils.git diff --git a/executors.py b/executors.py index 63efd81..6723bb9 100644 --- a/executors.py +++ b/executors.py @@ -17,6 +17,7 @@ import time from typing import Any, Callable, Dict, List, Optional, Set import cloudpickle # type: ignore +from overrides import overrides from ansi import bg, fg, underline, reset import argparse_utils @@ -64,7 +65,7 @@ SSH = 'ssh -oForwardX11=no' def make_cloud_pickle(fun, *args, **kwargs): - logger.info(f"Making cloudpickled bundle at {fun.__name__}") + logger.debug(f"Making cloudpickled bundle at {fun.__name__}") return cloudpickle.dumps((fun, args, kwargs)) @@ -121,6 +122,7 @@ class ThreadExecutor(BaseExecutor): self.histogram.add_item(duration) return result + @overrides def submit(self, function: Callable, *args, @@ -135,6 +137,7 @@ class ThreadExecutor(BaseExecutor): *newargs, **kwargs) + @overrides def shutdown(self, wait = True) -> None: logger.debug(f'Shutting down threadpool executor {self.title}') @@ -163,6 +166,7 @@ class ProcessExecutor(BaseExecutor): self.adjust_task_count(-1) return result + @overrides def submit(self, function: Callable, *args, @@ -181,6 +185,7 @@ class ProcessExecutor(BaseExecutor): ) return result + @overrides def shutdown(self, wait=True) -> None: logger.debug(f'Shutting down processpool executor {self.title}') self._process_executor.shutdown(wait) @@ -210,6 +215,7 @@ class RemoteWorkerRecord: class BundleDetails: pickled_code: bytes uuid: str + fname: str worker: Optional[RemoteWorkerRecord] username: Optional[str] machine: Optional[str] @@ -475,6 +481,7 @@ class RemoteExecutor(BaseExecutor): policy: RemoteWorkerSelectionPolicy) -> None: super().__init__() self.workers = workers + self.policy = policy self.worker_count = 0 for worker in self.workers: self.worker_count += worker.count @@ -482,7 +489,6 @@ class RemoteExecutor(BaseExecutor): msg = f"We need somewhere to schedule work; count was {self.worker_count}" logger.critical(msg) raise Exception(msg) - self.policy = policy self.policy.register_worker_pool(self.workers) self.cv = threading.Condition() self._helper_executor = fut.ThreadPoolExecutor( @@ -492,7 +498,7 @@ class RemoteExecutor(BaseExecutor): self.status = RemoteExecutorStatus(self.worker_count) self.total_bundles_submitted = 0 logger.debug( - f'Creating remote processpool with {self.worker_count} remote endpoints.' + f'Creating remote processpool with {self.worker_count} remote worker threads.' ) def is_worker_available(self) -> bool: @@ -586,6 +592,7 @@ class RemoteExecutor(BaseExecutor): bundle.worker = worker machine = bundle.machine = worker.machine username = bundle.username = worker.username + fname = bundle.fname self.status.record_acquire_worker(worker, uuid) logger.debug(f'Running bundle {uuid} on {worker}...') @@ -595,31 +602,31 @@ class RemoteExecutor(BaseExecutor): return self.post_launch_work(bundle) except Exception as e: logger.exception(e) - logger.info(f"Bundle {uuid} seems to have failed?!") + logger.info(f"{uuid}/{fname}: bundle seems to have failed?!") if bundle.failure_count < config.config['executors_max_bundle_failures']: return self.launch(bundle) else: - logger.info(f"Bundle {uuid} is poison, giving up on it.") + logger.info(f"{uuid}/{fname}: bundle is poison, giving up on it.") return None # Send input to machine if it's not local. if hostname not in machine: cmd = f'{RSYNC} {bundle.code_file} {username}@{machine}:{bundle.code_file}' - logger.info(f"Copying work to {worker} via {cmd}") + logger.info(f"{uuid}/{fname}: Copying work to {worker} via {cmd}") run_silently(cmd) # Do it. cmd = (f'{SSH} {bundle.username}@{bundle.machine} ' - f'"source remote-execution/bin/activate &&' + f'"source py39-venv/bin/activate &&' f' /home/scott/lib/python_modules/remote_worker.py' f' --code_file {bundle.code_file} --result_file {bundle.result_file}"') p = cmd_in_background(cmd, silent=True) bundle.pid = pid = p.pid - logger.info(f"Running {cmd} in the background as process {pid}") + logger.info(f"{uuid}/{fname}: Start training on {worker} via {cmd} (background pid {pid})") while True: try: - p.wait(timeout=0.5) + p.wait(timeout=0.25) except subprocess.TimeoutExpired: self.heartbeat() @@ -630,7 +637,7 @@ class RemoteExecutor(BaseExecutor): break else: logger.debug( - f"{pid}/{bundle.uuid} has finished its work normally." + f"{uuid}/{fname}: pid {pid} has finished its work normally." ) break @@ -638,10 +645,10 @@ class RemoteExecutor(BaseExecutor): return self.post_launch_work(bundle) except Exception as e: logger.exception(e) - logger.info(f"Bundle {uuid} seems to have failed?!") + logger.info(f"{uuid}: Bundle seems to have failed?!") if bundle.failure_count < config.config['executors_max_bundle_failures']: return self.launch(bundle) - logger.info(f"Bundle {uuid} is poison, giving up on it.") + logger.info(f"{uuid}: Bundle is poison, giving up on it.") return None def post_launch_work(self, bundle: BundleDetails) -> Any: @@ -652,6 +659,8 @@ class RemoteExecutor(BaseExecutor): machine = bundle.machine result_file = bundle.result_file code_file = bundle.code_file + fname = bundle.fname + uuid = bundle.uuid # Whether original or backup, if we finished first we must # fetch the results if the computation happened on a @@ -662,7 +671,7 @@ class RemoteExecutor(BaseExecutor): if bundle.hostname not in bundle.machine: cmd = f'{RSYNC} {username}@{machine}:{result_file} {result_file} 2>/dev/null' logger.info( - f"Fetching results from {username}@{machine} via {cmd}" + f"{uuid}/{fname}: Fetching results from {username}@{machine} via {cmd}" ) try: run_silently(cmd) @@ -686,7 +695,7 @@ class RemoteExecutor(BaseExecutor): # if one of the backups finished first; it still must read the # result from disk. if is_original: - logger.debug(f"Unpickling {result_file}.") + logger.debug(f"{uuid}/{fname}: Unpickling {result_file}.") try: with open(f'{result_file}', 'rb') as rb: serialized = rb.read() @@ -707,7 +716,7 @@ class RemoteExecutor(BaseExecutor): if bundle.backup_bundles is not None: for backup in bundle.backup_bundles: logger.debug( - f'Notifying backup {backup.uuid} that it is cancelled' + f'{uuid}/{fname}: Notifying backup {backup.uuid} that it\'s cancelled' ) backup.is_cancelled.set() @@ -721,7 +730,7 @@ class RemoteExecutor(BaseExecutor): # Tell the original to stop if we finished first. if not was_cancelled: logger.debug( - f'Notifying original {bundle.src_bundle.uuid} that it is cancelled' + f'{uuid}/{fname}: Notifying original {bundle.src_bundle.uuid} that it\'s cancelled' ) bundle.src_bundle.is_cancelled.set() @@ -730,7 +739,7 @@ class RemoteExecutor(BaseExecutor): self.adjust_task_count(-1) return result - def create_original_bundle(self, pickle): + def create_original_bundle(self, pickle, fname: str): from string_utils import generate_uuid uuid = generate_uuid(as_hex=True) code_file = f'/tmp/{uuid}.code.bin' @@ -743,6 +752,7 @@ class RemoteExecutor(BaseExecutor): bundle = BundleDetails( pickled_code = pickle, uuid = uuid, + fname = fname, worker = None, username = None, machine = None, @@ -761,7 +771,7 @@ class RemoteExecutor(BaseExecutor): failure_count = 0, ) self.status.record_bundle_details(bundle) - logger.debug(f'Created original bundle {uuid}') + logger.debug(f'{uuid}/{fname}: Created original bundle') return bundle def create_backup_bundle(self, src_bundle: BundleDetails): @@ -772,6 +782,7 @@ class RemoteExecutor(BaseExecutor): backup_bundle = BundleDetails( pickled_code = src_bundle.pickled_code, uuid = uuid, + fname = src_bundle.fname, worker = None, username = None, machine = None, @@ -791,7 +802,7 @@ class RemoteExecutor(BaseExecutor): ) src_bundle.backup_bundles.append(backup_bundle) self.status.record_bundle_details_already_locked(backup_bundle) - logger.debug(f'Created backup bundle {uuid}') + logger.debug(f'{uuid}/{src_bundle.fname}: Created backup bundle') return backup_bundle def schedule_backup_for_bundle(self, @@ -799,7 +810,7 @@ class RemoteExecutor(BaseExecutor): assert self.status.lock.locked() backup_bundle = self.create_backup_bundle(src_bundle) logger.debug( - f'Scheduling backup bundle {backup_bundle.uuid} for execution' + f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...' ) self._helper_executor.submit(self.launch, backup_bundle) @@ -807,15 +818,17 @@ class RemoteExecutor(BaseExecutor): # they will move the result_file to this machine and let # the original pick them up and unpickle them. + @overrides def submit(self, function: Callable, *args, **kwargs) -> fut.Future: pickle = make_cloud_pickle(function, *args, **kwargs) - bundle = self.create_original_bundle(pickle) + 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: self._helper_executor.shutdown(wait) logging.debug(f'Shutting down RemoteExecutor {self.title}') @@ -830,6 +843,7 @@ class DefaultExecutors(object): self.remote_executor: Optional[RemoteExecutor] = None def ping(self, host) -> bool: + logger.debug(f'RUN> ping -c 1 {host}') command = ['ping', '-c', '1', host] return subprocess.call( command, @@ -848,9 +862,11 @@ class DefaultExecutors(object): return self.process_executor def remote_pool(self) -> RemoteExecutor: + logger.info('Looking for some helper machines...') if self.remote_executor is None: pool: List[RemoteWorkerRecord] = [] if self.ping('cheetah.house'): + logger.info('Found cheetah.house') pool.append( RemoteWorkerRecord( username = 'scott', @@ -860,6 +876,7 @@ class DefaultExecutors(object): ), ) if self.ping('video.house'): + logger.info('Found video.house') pool.append( RemoteWorkerRecord( username = 'scott', @@ -869,6 +886,7 @@ class DefaultExecutors(object): ), ) if self.ping('wannabe.house'): + logger.info('Found wannabe.house') pool.append( RemoteWorkerRecord( username = 'scott', @@ -878,15 +896,17 @@ class DefaultExecutors(object): ), ) if self.ping('meerkat.cabin'): + logger.info('Found meerkat.cabin') pool.append( RemoteWorkerRecord( username = 'scott', machine = 'meerkat.cabin', - weight = 7, + weight = 5, count = 2, ), ) if self.ping('backup.house'): + logger.info('Found backup.house') pool.append( RemoteWorkerRecord( username = 'scott', @@ -895,7 +915,18 @@ class DefaultExecutors(object): count = 4, ), ) + if self.ping('kiosk.house'): + logger.info('Found kiosk.house') + pool.append( + RemoteWorkerRecord( + username = 'pi', + machine = 'kiosk.house', + weight = 1, + count = 2, + ), + ) if self.ping('puma.cabin'): + logger.info('Found puma.cabin') pool.append( RemoteWorkerRecord( username = 'scott', @@ -904,6 +935,13 @@ class DefaultExecutors(object): count = 4, ), ) + + # The controller machine has a lot to do; go easy on it. + for record in pool: + if record.machine == platform.node() and record.count > 1: + logger.info(f'Reducing workload for {record.machine}.') + record.count = 1 + policy = WeightedRandomRemoteWorkerSelectionPolicy() policy.register_worker_pool(pool) self.remote_executor = RemoteExecutor(pool, policy)