def _check_and_monitor_job(self, job_heartbeat_address, ping_heartbeat_address, max_memory): """ Sometimes the client may receive a job that is dead, thus we have to check if this job is still alive before adding it to the `actor_num`. """ # job_heartbeat_socket: sends heartbeat signal to job job_heartbeat_socket = self.ctx.socket(zmq.REQ) job_heartbeat_socket.linger = 0 job_heartbeat_socket.setsockopt(zmq.RCVTIMEO, int(0.9 * 1000)) job_heartbeat_socket.connect("tcp://" + ping_heartbeat_address) try: job_heartbeat_socket.send_multipart( [remote_constants.HEARTBEAT_TAG, to_byte(str(max_memory))]) job_heartbeat_socket.recv_multipart() except zmq.error.Again: job_heartbeat_socket.close(0) logger.error( "[Client] connects to a finished job, will try again, ping_heartbeat_address:{}" .format(ping_heartbeat_address)) return False job_heartbeat_socket.disconnect("tcp://" + ping_heartbeat_address) job_heartbeat_socket.connect("tcp://" + job_heartbeat_address) job_heartbeat_socket.setsockopt( zmq.RCVTIMEO, remote_constants.HEARTBEAT_TIMEOUT_S * 1000) # a thread for sending heartbeat signals to job thread = threading.Thread(target=self._create_job_monitor, args=(job_heartbeat_socket, )) thread.setDaemon(True) thread.start() return True
def send_file(self, socket): try: socket.send_multipart([ remote_constants.SEND_FILE_TAG, self.GLOBAL_CLIENT.pyfiles ]) _ = socket.recv_multipart() except zmq.error.Again as e: logger.error("Send python files failed.")
def wait_for_connection(self, reply_socket): """Wait for connection from the remote object. The remote object will send its class information and initialization arguments to the job, these parameters are then used to create a local instance in the job process. Args: reply_socket (sockert): main socket to accept commands of remote object. Returns: A local instance of the remote class object. """ message = reply_socket.recv_multipart() tag = message[0] obj = None if tag == remote_constants.INIT_OBJECT_TAG: try: file_name, class_name, end_of_file = cloudpickle.loads( message[1]) #/home/nlp-ol/Firework/baidu/nlp/evokit/python_api/es_agent -> es_agent file_name = file_name.split(os.sep)[-1] cls = load_remote_class(file_name, class_name, end_of_file) args, kwargs = cloudpickle.loads(message[2]) logfile_path = os.path.join(self.log_dir, 'stdout.log') with redirect_stdout_to_file(logfile_path): obj = cls(*args, **kwargs) except Exception as e: traceback_str = str(traceback.format_exc()) error_str = str(e) logger.error("traceback:\n{}".format(traceback_str)) reply_socket.send_multipart([ remote_constants.EXCEPTION_TAG, to_byte(error_str + "\ntraceback:\n" + traceback_str) ]) return None reply_socket.send_multipart([remote_constants.NORMAL_TAG]) else: logger.error("Message from job {}".format(message)) reply_socket.send_multipart([ remote_constants.EXCEPTION_TAG, b"[job]Unkonwn tag when tried to receive the class definition" ]) raise NotImplementedError return obj
def test(): # 创建环境 game = FlappyBird() env = PLE(game, fps=30, display_screen=True) obs_dim = len(env.getGameState()) act_dim = len(env.getActionSet()) print('action set:', env.getActionSet()) logger.info('obs_dim {}, act_dim {}'.format(obs_dim, act_dim)) # 创建经验池 rpm = ReplayMemory(MEMORY_SIZE) # DQN的经验回放池 # 根据parl框架构建agent model = Model(act_dim=act_dim) algorithm = DQN(model, act_dim=act_dim, gamma=GAMMA, lr=LEARNING_RATE) agent = Agent(algorithm, obs_dim=obs_dim, act_dim=act_dim, e_greed=0.3, e_greed_decrement=1e-6) # 加载模型 save_path = './DQN/checkpoints/episode_V14600.ckpt' print('checkpoints:', save_path) if os.path.exists(save_path): logger.info('load ckpt success!') agent.restore(save_path) else: logger.error('load ckpt error!') action_set = env.getActionSet() env.init() episode_reward = 0 steps = 0 while not env.game_over(): steps += 1 if (steps == 1): continue obs = list(env.getGameState().values()) action_idx = agent.predict(obs) # 预测动作,只选最优动作 act = action_set[action_idx] reward = env.act(act) episode_reward += reward reward_str = str(int(episode_reward)) drawText(env.game.screen, reward_str, 288, 0, 48, (255, 0, 0), (255, 255, 255)) env.reset_game() logger.info('[Test] steps:{}, reward:{}'.format(steps, episode_reward))
def wait_for_files(self, reply_socket, job_address): """Wait for python files from remote object. When a remote object receives the allocated job address, it will send the python files to the job. Later, the job will save these files to a temporary directory and add the temporary diretory to Python's working directory. Args: reply_socket (sockert): main socket to accept commands of remote object. job_address (String): address of reply_socket. Returns: A temporary directory containing the python files. """ message = reply_socket.recv_multipart() tag = message[0] if tag == remote_constants.SEND_FILE_TAG: pyfiles = pickle.loads(message[1]) # save python files to temporary directory envdir = tempfile.mkdtemp() for file, code in pyfiles['python_files'].items(): file = os.path.join(envdir, file) with open(file, 'wb') as code_file: code_file.write(code) # save other files to current directory for file, content in pyfiles['other_files'].items(): # create directory (i.e. ./rom_files/) if '/' in file: try: sep = os.sep recursive_dirs = os.path.join(*(file.split(sep)[:-1])) recursive_dirs = os.path.join(envdir, recursive_dirs) os.makedirs(recursive_dirs) except OSError as e: pass file = os.path.join(envdir, file) with open(file, 'wb') as f: f.write(content) reply_socket.send_multipart([remote_constants.NORMAL_TAG]) return envdir else: logger.error("NotImplementedError:{}, received tag:{}".format( job_address, )) raise NotImplementedError
def run(self, job_address_sender, job_id_sender): """An infinite loop waiting for a new task. Args: job_address_sender(sending end of multiprocessing.Pipe): send job address of reply_socket to main process. """ ctx = zmq.Context() # create the reply_socket reply_socket = ctx.socket(zmq.REP) job_port = reply_socket.bind_to_random_port(addr="tcp://*") reply_socket.linger = 0 job_ip = get_ip_address() job_address = "{}:{}".format(job_ip, job_port) job_id = job_address.replace(':', '_') + '_' + str(int(time.time())) self.log_dir = os.path.expanduser('~/.parl_data/job/{}'.format(job_id)) logger.set_dir(self.log_dir) logger.info( "[Job] Job {} initialized. Reply heartbeat socket Address: {}.". format(job_id, job_address)) job_address_sender.send(job_address) job_id_sender.send(job_id) try: # receive source code from the actor and append them to the environment variables. envdir = self.wait_for_files(reply_socket, job_address) sys.path.insert(0, envdir) os.chdir(envdir) obj = self.wait_for_connection(reply_socket) assert obj is not None self.single_task(obj, reply_socket, job_address) except Exception as e: logger.error( "Error occurs when running a single task. We will reset this job. \nReason:{}" .format(e)) traceback_str = str(traceback.format_exc()) logger.error("traceback:\n{}".format(traceback_str))
def playGame(self, verbose=False): """ Executes one episode of a game. Returns: either winner: player who won the game (1 if player1, -1 if player2) or draw result returned from the game that is neither 1, -1, nor 0. """ players = [self.player2, None, self.player1] curPlayer = 1 board = self.game.getInitBoard() it = 0 while self.game.getGameEnded(board, curPlayer) == 0: it += 1 if verbose: assert self.display print("Turn ", str(it), "Player ", str(curPlayer)) self.display(board) action = players[curPlayer + 1](self.game.getCanonicalForm( board, curPlayer)) valids = self.game.getValidMoves( self.game.getCanonicalForm(board, curPlayer), 1) if valids[action] == 0: logger.error('Action {} is not valid!'.format(action)) logger.debug('valids = {}'.format(valids)) assert valids[action] > 0 board, curPlayer = self.game.getNextState(board, curPlayer, action) if verbose: assert self.display print("Game over: Turn ", str(it), "Result ", str(self.game.getGameEnded(board, 1))) self.display(board) return curPlayer * self.game.getGameEnded(board, curPlayer)
def _reply_client_heartbeat(self, socket): """Create a socket that replies heartbeat signals from the client. If the job losts connection with the client, it will exit too. """ while True: try: message = socket.recv_multipart() stop_job = self._check_used_memory() socket.send_multipart([ remote_constants.HEARTBEAT_TAG, to_byte(str(stop_job)), to_byte(self.job_address) ]) if stop_job == True: logger.error( "Memory used by this job exceeds {}. This job will exist." .format(self.max_memory)) time.sleep(5) socket.close(0) os._exit(1) except zmq.error.Again as e: logger.warning( "[Job] Cannot connect to the client. This job will exit and inform the worker." ) break socket.close(0) with self.lock: self.kill_job_socket.send_multipart( [remote_constants.KILLJOB_TAG, to_byte(self.job_address)]) try: _ = self.kill_job_socket.recv_multipart() except zmq.error.Again as e: pass logger.warning("[Job]lost connection with the client, will exit") os._exit(1)
def _create_job_monitor(self, job_heartbeat_socket): """Send heartbeat signals to check target's status""" job_is_alive = True while job_is_alive and self.client_is_alive: try: job_heartbeat_socket.send_multipart( [remote_constants.HEARTBEAT_TAG]) job_message = job_heartbeat_socket.recv_multipart() stop_job = to_str(job_message[1]) job_address = to_str(job_message[2]) if stop_job == 'True': logger.error( 'Job {} exceeds max memory usage, will stop this job.'. format(job_address)) self.lock.acquire() self.actor_num -= 1 self.lock.release() job_is_alive = False else: time.sleep(remote_constants.HEARTBEAT_INTERVAL_S) except zmq.error.Again as e: job_is_alive = False self.lock.acquire() self.actor_num -= 1 logger.error( '[xparl] lost connection with a job, current actor num: {}' .format(self.actor_num)) self.lock.release() except zmq.error.ZMQError as e: break job_heartbeat_socket.close(0)
def _create_jobs(self): """Create jobs and send a instance of ``InitializedWorker`` that contains the worker information to the master.""" try: self.request_master_socket.send_multipart( [remote_constants.WORKER_CONNECT_TAG]) _ = self.request_master_socket.recv_multipart() except zmq.error.Again as e: logger.error("Can not connect to the master, " "please check if master is started.") self.master_is_alive = False return initialized_jobs = self._init_jobs(job_num=self.cpu_num) self.request_master_socket.setsockopt( zmq.RCVTIMEO, remote_constants.HEARTBEAT_TIMEOUT_S * 1000) self.reply_master_hearbeat_thread = threading.Thread( target=self._reply_heartbeat, args=("master {}".format(self.master_address), )) self.reply_master_hearbeat_thread.start() self.heartbeat_socket_initialized.wait() for job in initialized_jobs: job.worker_address = self.master_heartbeat_address initialized_worker = InitializedWorker(self.master_heartbeat_address, initialized_jobs, self.cpu_num, socket.gethostname()) self.request_master_socket.send_multipart([ remote_constants.WORKER_INITIALIZED_TAG, cloudpickle.dumps(initialized_worker) ]) _ = self.request_master_socket.recv_multipart() self.worker_status = WorkerStatus(self.master_heartbeat_address, initialized_jobs, self.cpu_num)
def single_task(self, obj, reply_socket, job_address): """An infinite loop waiting for commands from the remote object. Each job will receive two kinds of message from the remote object: 1. When the remote object calls a function, job will run the function on the local instance and return the results to the remote object. 2. When the remote object is deleted, the job will quit and release related computation resources. Args: reply_socket (sockert): main socket to accept commands of remote object. job_address (String): address of reply_socket. """ while True: message = reply_socket.recv_multipart() tag = message[0] if tag in [ remote_constants.CALL_TAG, remote_constants.GET_ATTRIBUTE, remote_constants.SET_ATTRIBUTE ]: # if tag == remote_constants.CALL_TAG: try: if tag == remote_constants.CALL_TAG: function_name = to_str(message[1]) data = message[2] args, kwargs = loads_argument(data) # Redirect stdout to stdout.log temporarily logfile_path = os.path.join(self.log_dir, 'stdout.log') with redirect_stdout_to_file(logfile_path): ret = getattr(obj, function_name)(*args, **kwargs) ret = dumps_return(ret) reply_socket.send_multipart( [remote_constants.NORMAL_TAG, ret]) elif tag == remote_constants.GET_ATTRIBUTE: attribute_name = to_str(message[1]) logfile_path = os.path.join(self.log_dir, 'stdout.log') with redirect_stdout_to_file(logfile_path): ret = getattr(obj, attribute_name) ret = dumps_return(ret) reply_socket.send_multipart( [remote_constants.NORMAL_TAG, ret]) else: attribute_name = to_str(message[1]) attribute_value = loads_return(message[2]) logfile_path = os.path.join(self.log_dir, 'stdout.log') with redirect_stdout_to_file(logfile_path): setattr(obj, attribute_name, attribute_value) reply_socket.send_multipart( [remote_constants.NORMAL_TAG]) except Exception as e: # reset the job error_str = str(e) logger.error(error_str) if type(e) == AttributeError: reply_socket.send_multipart([ remote_constants.ATTRIBUTE_EXCEPTION_TAG, to_byte(error_str) ]) raise AttributeError elif type(e) == SerializeError: reply_socket.send_multipart([ remote_constants.SERIALIZE_EXCEPTION_TAG, to_byte(error_str) ]) raise SerializeError elif type(e) == DeserializeError: reply_socket.send_multipart([ remote_constants.DESERIALIZE_EXCEPTION_TAG, to_byte(error_str) ]) raise DeserializeError else: traceback_str = str(traceback.format_exc()) logger.error("traceback:\n{}".format(traceback_str)) reply_socket.send_multipart([ remote_constants.EXCEPTION_TAG, to_byte(error_str + "\ntraceback:\n" + traceback_str) ]) break # receive DELETE_TAG from actor, and stop replying worker heartbeat elif tag == remote_constants.KILLJOB_TAG: reply_socket.send_multipart([remote_constants.NORMAL_TAG]) logger.warning( "An actor exits and this job {} will exit.".format( job_address)) break else: logger.error( "The job receives an unknown message: {}".format(message)) raise NotImplementedError