X-Git-Url: https://wannabe.guru.org/gitweb/?a=blobdiff_plain;f=executors.py;h=c11bd546cc3b1b0afcefff8c3215aab2707db147;hb=ed8fa2b10b0177b15b7423263bdd390efde2f0c8;hp=b446822945a73c7a896755fc4b509d24c3600f84;hpb=3f4818bf9d1f3071c4e2906896d422810b5776bc;p=python_utils.git diff --git a/executors.py b/executors.py index b446822..c11bd54 100644 --- a/executors.py +++ b/executors.py @@ -22,8 +22,8 @@ from overrides import overrides from ansi import bg, fg, underline, reset import argparse_utils import config -from exec_utils import run_silently, cmd_in_background, cmd_with_timeout from decorator_utils import singleton +from exec_utils import run_silently, cmd_in_background, cmd_with_timeout import histogram as hist logger = logging.getLogger(__name__) @@ -230,8 +230,8 @@ class BundleDetails: pid: int start_ts: float end_ts: float - too_slow: bool - super_slow: bool + slower_than_local_p95: bool + slower_than_global_p95: bool src_bundle: BundleDetails is_cancelled: threading.Event was_cancelled: bool @@ -419,21 +419,19 @@ class RemoteExecutorStatus: if qworker is not None: if sec > qworker[1]: ret += f'{bg("red")}>💻p95{reset()} ' - elif sec > qworker[0]: - ret += f'{fg("red")}>💻p50{reset()} ' - if qall is not None: - if sec > qall[1] * 1.5: - ret += f'{bg("red")}!!!{reset()}' if details is not None: - logger.debug(f'Flagging {details} for another backup') - details.super_slow = True - elif sec > qall[1]: + details.slower_than_local_p95 = True + else: + if details is not None: + details.slower_than_local_p95 = False + + if qall is not None: + if sec > qall[1]: ret += f'{bg("red")}>∀p95{reset()} ' if details is not None: - logger.debug(f'Flagging {details} for a backup') - details.too_slow = True - elif sec > qall[0]: - ret += f'{fg("red")}>∀p50{reset()}' + details.slower_than_global_p95 = True + else: + details.slower_than_global_p95 = False ret += '\n' return ret @@ -451,7 +449,6 @@ class RemoteExecutorStatus: class RemoteWorkerSelectionPolicy(ABC): def register_worker_pool(self, workers): - random.seed() self.workers = workers @abstractmethod @@ -467,12 +464,14 @@ class RemoteWorkerSelectionPolicy(ABC): class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): + @overrides def is_worker_available(self) -> bool: for worker in self.workers: if worker.count > 0: return True return False + @overrides def acquire_worker( self, machine_to_avoid = None @@ -499,12 +498,14 @@ class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): def __init__(self) -> None: self.index = 0 + @overrides def is_worker_available(self) -> bool: for worker in self.workers: if worker.count > 0: return True return False + @overrides def acquire_worker( self, machine_to_avoid: str = None @@ -551,6 +552,8 @@ class RemoteExecutor(BaseExecutor): ) self.status = RemoteExecutorStatus(self.worker_count) self.total_bundles_submitted = 0 + self.backup_lock = threading.Lock() + self.last_backup = None def is_worker_available(self) -> bool: return self.policy.is_worker_available() @@ -588,37 +591,84 @@ class RemoteExecutor(BaseExecutor): # Look for bundles to reschedule. num_done = len(self.status.finished_bundle_timings) - if num_done > 7 or (num_done > 5 and self.is_worker_available()): - for worker, bundle_uuids in self.status.in_flight_bundles_by_worker.items(): - for uuid in bundle_uuids: - bundle = self.status.bundle_details_by_uuid.get(uuid, None) - if ( - bundle is not None and - bundle.too_slow and - bundle.src_bundle is None and - config.config['executors_schedule_remote_backups'] - ): - self.consider_backup_for_bundle(bundle) - - def consider_backup_for_bundle(self, bundle: BundleDetails) -> None: - assert self.status.lock.locked() - if ( - bundle.too_slow - and len(bundle.backup_bundles) == 0 # one backup per - ): - msg = f"*** Rescheduling {bundle} (first backup) ***" - logger.debug(msg) - self.schedule_backup_for_bundle(bundle) - return - elif ( - bundle.super_slow - and len(bundle.backup_bundles) < 2 # two backups in dire situations - and self.status.total_idle() > 4 - ): - msg = f"*** Rescheduling {bundle} (second backup) ***" - logger.debug(msg) - self.schedule_backup_for_bundle(bundle) - return + num_idle_workers = self.worker_count - self.task_count + now = time.time() + if ( + config.config['executors_schedule_remote_backups'] + and num_done > 2 + and num_idle_workers > 1 + and (self.last_backup is None or (now - self.last_backup > 1.0)) + and self.backup_lock.acquire(blocking=False) + ): + try: + assert self.backup_lock.locked() + + bundle_to_backup = None + best_score = None + for worker, bundle_uuids in self.status.in_flight_bundles_by_worker.items(): + # Prefer to schedule backups of bundles on slower machines. + base_score = 0 + for record in self.workers: + if worker.machine == record.machine: + base_score = float(record.weight) + base_score = 1.0 / base_score + base_score *= 200.0 + base_score = int(base_score) + break + + for uuid in bundle_uuids: + bundle = self.status.bundle_details_by_uuid.get(uuid, None) + if ( + bundle is not None + and bundle.src_bundle is None + and bundle.backup_bundles is not None + ): + score = base_score + + # Schedule backups of bundles running longer; especially those + # that are unexpectedly slow. + start_ts = self.status.start_per_bundle[uuid] + if start_ts is not None: + runtime = now - start_ts + score += runtime + logger.debug(f'score[{bundle}] => {score} # latency boost') + + if bundle.slower_than_local_p95: + score += runtime / 2 + logger.debug(f'score[{bundle}] => {score} # >worker p95') + + if bundle.slower_than_global_p95: + score += runtime / 2 + logger.debug(f'score[{bundle}] => {score} # >global p95') + + # Prefer backups of bundles that don't have backups already. + backup_count = len(bundle.backup_bundles) + if backup_count == 0: + score *= 2 + elif backup_count == 1: + score /= 2 + elif backup_count == 2: + score /= 8 + else: + score = 0 + logger.debug(f'score[{bundle}] => {score} # {backup_count} dup backup factor') + + if ( + score != 0 + and (best_score is None or score > best_score) + ): + bundle_to_backup = bundle + assert bundle is not None + assert bundle.backup_bundles is not None + assert bundle.src_bundle is None + best_score = score + + if bundle_to_backup is not None: + self.last_backup = now + logger.info(f'=====> SCHEDULING BACKUP {bundle_to_backup} (score={best_score:.1f}) <=====') + self.schedule_backup_for_bundle(bundle_to_backup) + finally: + self.backup_lock.release() def check_if_cancelled(self, bundle: BundleDetails) -> bool: with self.status.lock: @@ -921,8 +971,8 @@ class RemoteExecutor(BaseExecutor): pid = 0, start_ts = time.time(), end_ts = 0.0, - too_slow = False, - super_slow = False, + slower_than_local_p95 = False, + slower_than_global_p95 = False, src_bundle = None, is_cancelled = threading.Event(), was_cancelled = False, @@ -951,8 +1001,8 @@ class RemoteExecutor(BaseExecutor): pid = 0, start_ts = time.time(), end_ts = 0.0, - too_slow = False, - super_slow = False, + slower_than_local_p95 = False, + slower_than_global_p95 = False, src_bundle = src_bundle, is_cancelled = threading.Event(), was_cancelled = False, @@ -967,6 +1017,7 @@ class RemoteExecutor(BaseExecutor): def schedule_backup_for_bundle(self, src_bundle: BundleDetails): assert self.status.lock.locked() + assert src_bundle is not None backup_bundle = self.create_backup_bundle(src_bundle) logger.debug( f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...' @@ -1095,16 +1146,6 @@ class DefaultExecutors(object): count = 2, ), ) - if self.ping('backup.house'): - logger.info('Found backup.house') - pool.append( - RemoteWorkerRecord( - username = 'scott', - machine = 'backup.house', - weight = 1, - count = 4, - ), - ) if self.ping('kiosk.house'): logger.info('Found kiosk.house') pool.append(