#!/usr/bin/env python # coding utf-8 import asyncio import json import socket as pysocket from logging import getLogger from random import randint from time import perf_counter import numpy as np import zmq from monty.json import jsanitize from monty.serialization import MontyDecoder from maggma.cli.multiprocessing import multi from maggma.cli.settings import CLISettings from maggma.core import Builder from maggma.utils import tqdm settings = CLISettings() def find_port(): sock = pysocket.socket() sock.bind(("", 0)) return sock.getsockname()[1] def manager(url: str, port: int, builders: list[Builder], num_chunks: int, num_workers: int): """ Really simple manager for distributed processing that uses a builder prechunk to modify the builder and send out modified builders for each worker to run. The manager will try and keep track of workers, including which error out and which complete. Currently, if a single workers fails the entire distributed job will be stopped. """ logger = getLogger("Manager") if not (num_chunks and num_workers): raise ValueError("Both num_chunks and num_workers must be non-zero") logger.info(f"Binding to Manager URL {url}:{port}") # Setup socket and polling socket, poll = setup(url, port) workers = {} # type: ignore logger.debug("Manager started and looking for workers") for builder in builders: logger.info(f"Working on {builder.__class__.__name__}") builder_dict = builder.as_dict() try: builder.connect() chunk_dicts = [{"chunk": d, "distributed": False, "completed": False} for d in builder.prechunk(num_chunks)] pbar_distributed = tqdm( total=len(chunk_dicts), desc=f"Distributed chunks for {builder.__class__.__name__}", ) pbar_completed = tqdm( total=len(chunk_dicts), desc=f"Completed chunks for {builder.__class__.__name__}", ) logger.info(f"Distributing {len(chunk_dicts)} chunks to workers") except NotImplementedError: attempt_graceful_shutdown(workers, socket) raise RuntimeError(f"Can't distribute process {builder.__class__.__name__} as no prechunk method exists.") completed = False while not completed: completed = all(chunk["completed"] for chunk in chunk_dicts) if num_workers <= 0: socket.close() raise RuntimeError("No workers to distribute chunks to") # Poll and look for messages from workers connections = dict(poll.poll(100)) # If workers send messages decode and figure out what do if connections: identity, _, bmsg = socket.recv_multipart() msg = bmsg.decode("utf-8") if "READY" in msg: if identity not in workers: logger.debug(f"Got connection from worker: {msg.split('_')[1]}") workers[identity] = { "working": False, "heartbeats": 1, "last_ping": perf_counter(), "work_index": -1, } else: workers[identity]["working"] = False work_ind = workers[identity]["work_index"] if work_ind != -1: chunk_dicts[work_ind]["completed"] = True # type: ignore pbar_completed.update(1) # If everything is distributed, send EXIT to the worker if all(chunk["distributed"] for chunk in chunk_dicts): logger.debug(f"Sending exit signal to worker: {msg.split('_')[1]}") socket.send_multipart([identity, b"", b"EXIT"]) workers.pop(identity) elif "ERROR" in msg: # Remove worker and requeue work sent to it attempt_graceful_shutdown(workers, socket) raise RuntimeError( "At least one worker has stopped with error message: {}".format(msg.split("_")[1]) ) elif msg == "PING": # Respond to heartbeat socket.send_multipart([identity, b"", b"PONG"]) workers[identity]["last_ping"] = perf_counter() workers[identity]["heartbeats"] += 1 # Decide if any workers are dead and need to be removed if settings.WORKER_TIMEOUT is not None: handle_dead_workers(workers, socket) for work_index, chunk_dict in enumerate(chunk_dicts): if not chunk_dict["distributed"]: temp_builder_dict = dict(**builder_dict) temp_builder_dict.update(chunk_dict["chunk"]) # type: ignore temp_builder_dict = jsanitize(temp_builder_dict, recursive_msonable=True) # Send work for available workers for identity in workers: if not workers[identity]["working"]: # Send out a chunk to idle worker socket.send_multipart( [ identity, b"", json.dumps(temp_builder_dict).encode("utf-8"), ] ) workers[identity]["work_index"] = work_index workers[identity]["working"] = True chunk_dicts[work_index]["distributed"] = True pbar_distributed.update(1) # Send EXIT to any remaining workers logger.info("Sending exit messages to workers once they are done") attempt_graceful_shutdown(workers, socket) def setup(url, port): context = zmq.Context() context.setsockopt(opt=zmq.SocketOption.ROUTER_MANDATORY, value=1) context.setsockopt(opt=zmq.SNDHWM, value=0) context.setsockopt(opt=zmq.RCVHWM, value=0) socket = context.socket(zmq.ROUTER) socket.bind(f"{url}:{port}") poll = zmq.Poller() poll.register(socket, zmq.POLLIN) return socket, poll def attempt_graceful_shutdown(workers, socket): for identity in workers: socket.send_multipart([identity, b"", b"EXIT"]) socket.close() def handle_dead_workers(workers, socket): if len(workers) == 1: # Use global timeout identity = next(iter(workers.keys())) if (perf_counter() - workers[identity]["last_ping"]) >= settings.WORKER_TIMEOUT: attempt_graceful_shutdown(workers, socket) raise RuntimeError("Worker has timed out. Stopping distributed build.") elif len(workers) == 2: # Use 10% ratio between workers workers_sorted = sorted(workers.items(), key=lambda x: x[1]["heartbeats"]) ratio = workers_sorted[1][1]["heartbeats"] / workers_sorted[0][1]["heartbeats"] if ratio <= 0.1: attempt_graceful_shutdown(workers, socket) raise RuntimeError("One worker has timed out. Stopping distributed build.") elif len(workers) > 2: # Calculate modified z-score of heartbeat counts and see if any are <= -3.5 hearbeat_vals = [w["heartbeats"] for w in workers.values()] median = np.median(hearbeat_vals) mad = np.median([abs(i - median) for i in hearbeat_vals]) if mad > 0: for identity in list(workers.keys()): z_score = 0.6745 * (workers[identity]["heartbeats"] - median) / mad if z_score <= -3.5: attempt_graceful_shutdown(workers, socket) raise RuntimeError("At least one worker has timed out. Stopping distributed build.") def worker(url: str, port: int, num_processes: int, no_bars: bool): """ Simple distributed worker that connects to a manager asks for work and deploys using multiprocessing. """ identity = f"{randint(0, 0x10000):04X}-{randint(0, 0x10000):04X}" logger = getLogger(f"Worker {identity}") logger.info(f"Connecting to Manager at {url}:{port}") context = zmq.Context() socket: zmq.Socket = context.socket(zmq.REQ) socket.setsockopt_string(zmq.IDENTITY, identity) socket.connect(f"{url}:{port}") poller = zmq.Poller() poller.register(socket, zmq.POLLIN) # Initial message package hostname = pysocket.gethostname() try: running = True while running: socket.send(f"READY_{hostname}".encode()) # Poll for MANAGER_TIMEOUT seconds, if nothing is given then assume manager is dead and timeout connections = dict(poller.poll(settings.MANAGER_TIMEOUT * 1000)) if not connections: socket.close() raise RuntimeError("Stopping work as manager timed out.") bmessage: bytes = socket.recv() message = bmessage.decode("utf-8") if "@class" in message and "@module" in message: # We have a valid builder work = json.loads(message) builder = MontyDecoder().process_decoded(work) asyncio.run( multi( builder, num_processes, no_bars=no_bars, heartbeat_func=ping_manager, heartbeat_func_kwargs={"socket": socket, "poller": poller}, ) ) elif message == "EXIT": # End the worker running = False except Exception as e: logger.error(f"A worker failed with error: {e}") socket.send(f"ERROR_{e}".encode()) socket.close() socket.close() def ping_manager(socket, poller): socket.send_string("PING") # Poll for MANAGER_TIMEOUT seconds, if nothing is given then assume manager is dead and timeout connections = dict(poller.poll(settings.MANAGER_TIMEOUT * 1000)) if not connections: socket.close() raise RuntimeError("Stopping work as manager timed out.") message: bytes = socket.recv() if message.decode("utf-8") != "PONG": socket.close() raise RuntimeError("Stopping work as manager did not respond to heartbeat from worker.")