# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Defination of Role Makers.""" import os import time import numpy as np import warnings from multiprocessing import Process, Manager import paddle import paddle.fluid as fluid from paddle.distributed.fleet.base.private_helper_function import wait_server_ready class Role: WORKER = 1 SERVER = 2 HETER_WORKER = 3 ALL = 4 class Gloo(object): """ Gloo is a universal class for barrier and collective communication """ class RENDEZVOUS: HDFS = 1 FILE = 2 HTTP = 3 def __init__(self): self._worker_comm = None self._server_comm = None self._nodes_comm = None self._comm_world = ["worker", "server", "all"] self._err_init = "gloo is not initialized, will not communicator with other nodes" self._err_type = "gloo initialized error, please check arguments" self._err_world = "argument error, comm_world must in {}".format( self._comm_world) self._is_initialized = False self._init_timeout_seconds = 3600 self._run_timeout_seconds = 9999999 self._rendezvous = None self._role = None self._iface = None self._role_id = -1 self._worker_num = -1 self._server_num = -1 self._need_init_all = False def init(self, rendezvous, role, role_id, worker_num, server_num, need_init_all=False, kwargs=None): self._rendezvous = rendezvous self._role = role self._role_id = role_id self._worker_num = worker_num self._server_num = server_num self._need_init_all = need_init_all self._iface = "" self._prefix = kwargs.get("store.prefix", "") http_server = None if self._rendezvous == Gloo.RENDEZVOUS.HDFS: dfs_name = kwargs.get("dfs.name", "") dfs_ugi = kwargs.get("dfs.ugi", "") dfs_path = kwargs.get("dfs.path", "") if not dfs_name or not dfs_ugi or not dfs_path: raise ValueError(self._err_type) self._init_dfs(dfs_name, dfs_ugi, dfs_path, self._prefix) elif self._rendezvous == Gloo.RENDEZVOUS.FILE: fs_path = kwargs.get("dfs.path", "") if not fs_path: raise ValueError(self._err_type) self._init_fs(fs_path, self._prefix) elif self._rendezvous == Gloo.RENDEZVOUS.HTTP: ip = kwargs.get("http.host", "") port = kwargs.get("http.port", "") start_http_server = kwargs.get("start_http_server", False) http_server_d = kwargs.get("http_server_d") if not ip or not port: raise ValueError(self._err_type) http_server = self._init_http(ip, port, self._prefix, start_http_server, http_server_d) else: raise ValueError(self._err_type) self._is_initialized = True self._http_server = http_server def _init_fs(self, fs_path, prefix): def init(rank, nodes, role): gloo = fluid.core.Gloo() gloo.set_rank(rank) gloo.set_size(nodes) gloo.set_prefix(prefix) gloo.set_iface(self._iface) gloo.set_timeout_seconds(self._init_timeout_seconds, self._run_timeout_seconds) gloo.set_hdfs_store(os.path.join(fs_path, role), "", "") gloo.init() return gloo if self._role == Role.WORKER: rank, nodes = self._get_rank_nodes(Role.WORKER) gloo = init(rank, nodes, "WORKER") self._worker_comm = gloo else: rank, nodes = self._get_rank_nodes(Role.SERVER) gloo = init(rank, nodes, "SERVER") self._server_comm = gloo if self._need_init_all: rank, nodes = self._get_rank_nodes(Role.ALL) gloo = init(rank, nodes, "ALL") self._nodes_comm = gloo def _init_dfs(self, dfs_name, dfs_ugi, dfs_path, prefix): def init(rank, nodes, role): gloo = fluid.core.Gloo() gloo.set_rank(rank) gloo.set_size(nodes) gloo.set_prefix(prefix) gloo.set_iface(self._iface) gloo.set_timeout_seconds(self._init_timeout_seconds, self._run_timeout_seconds) gloo.set_hdfs_store(os.path.join(dfs_path, role), dfs_name, dfs_ugi) gloo.init() return gloo if self._role == Role.WORKER: rank, nodes = self._get_rank_nodes(Role.WORKER) gloo = init(rank, nodes, "WORKER") self._worker_comm = gloo else: rank, nodes = self._get_rank_nodes(Role.SERVER) gloo = init(rank, nodes, "SERVER") self._server_comm = gloo if self._need_init_all: rank, nodes = self._get_rank_nodes(Role.ALL) gloo = init(rank, nodes, "ALL") self._nodes_comm = gloo def _init_http(self, ip, port, prefix, start_http_server, http_server_d): def __start_kv_server(http_server_d, size_d): from paddle.distributed.fleet.utils.http_server import KVServer http_server = KVServer(port, size_d) http_server.start() wait_seconds = 5 while http_server_d.get("running", False): time.sleep(wait_seconds) http_server.stop() def init_kv_server(http_server_d): size_d = { "trainer": self._worker_num, "pserver": self._server_num, "all": self._worker_num + self._server_num } http_server_d["running"] = True # child process for http server _http_server = Process( target=__start_kv_server, args=(http_server_d, size_d)) _http_server.daemon = True # set running status to True # start child process _http_server.start() return _http_server def init(rank, nodes, role): gloo = fluid.core.Gloo() gloo.set_rank(rank) gloo.set_size(nodes) gloo.set_prefix(prefix) gloo.set_iface(self._iface) gloo.set_timeout_seconds(self._init_timeout_seconds, self._run_timeout_seconds) gloo.set_http_store(ip, port, role) ep = ":".join([ip, str(port)]) wait_server_ready([ep]) gloo.init() return gloo port = int(port) if start_http_server: http_server = init_kv_server(http_server_d) if self._role == Role.WORKER: rank, nodes = self._get_rank_nodes(Role.WORKER) gloo = init(rank, nodes, "WORKER") self._worker_comm = gloo else: rank, nodes = self._get_rank_nodes(Role.SERVER) gloo = init(rank, nodes, "SERVER") self._server_comm = gloo if self._need_init_all: rank, nodes = self._get_rank_nodes(Role.ALL) gloo = init(rank, nodes, "ALL") self._nodes_comm = gloo if start_http_server: http_server_d["running"] = False http_server.join() def _get_rank_nodes(self, role): nodes = 0 rank = -1 if role == Role.WORKER: nodes = self._worker_num rank = self._role_id elif role == Role.SERVER: nodes = self._server_num rank = self._role_id elif role == Role.ALL: nodes = self._worker_num + self._server_num if self._role == Role.WORKER: rank = self._role_id else: rank = self._worker_num + self._role_id else: ValueError(self._err_type) return rank, nodes def __get_default_iface(self): """ get default physical interface """ default1 = self.__get_default_iface_from_gateway() default2 = self.__get_default_iface_from_interfaces() return default2 if default1 == "lo" else default1 def __get_default_iface_from_gateway(self): """ get default physical interface """ res = os.popen("route -A inet").read().strip().split("\n") gateway_idx = None iface_idx = None for item in res: item = item.split() if "Gateway" in item and "Iface" in item: gateway_idx = item.index("Gateway") iface_idx = item.index("Iface") elif gateway_idx != None and iface_idx != None: gateway = None if len(item) > gateway_idx: gateway = item[gateway_idx] if gateway and gateway != '*' and gateway != "0.0.0.0" and len( item) > iface_idx: return item[iface_idx] return "lo" def __get_default_iface_from_interfaces(self): """ get default physical interface """ res = os.popen("ip -f inet addr | awk NR%3==1").read().strip().split( "\n") for item in res: if "BROADCAST" in item: return item.split(":")[1].strip() return "lo" def barrier(self, comm_world): """ dummy barrier, do nothing """ if not self._is_initialized: warnings.warn(self._err_init) return if comm_world not in self._comm_world: raise ValueError(self._err_world) if comm_world == "worker": self._worker_comm.barrier() elif comm_world == "server": self._server_comm.barrier() else: self._nodes_comm.barrier() def all_reduce(self, input, mode="sum", comm_world="worker"): if not self._is_initialized: warnings.warn(self._err_init) return input if comm_world not in self._comm_world: raise ValueError(self._err_world) input = np.array(input) input_shape = input.shape input_list = input.reshape(-1).tolist() self.barrier(comm_world) if comm_world == "worker": ans = self._worker_comm.all_reduce(input_list, mode) elif comm_world == "server": ans = self._server_comm.all_reduce(input_list, mode) else: ans = self._nodes_comm.all_reduce(input_list, mode) output = np.array(ans).reshape(input_shape) return output def all_gather(self, input, comm_world="worker"): """ dummy all gather, do nothing Args: obj(any): obj to do all gather """ if not self._is_initialized: warnings.warn(self._err_init) return input if comm_world not in self._comm_world: raise ValueError(self._err_world) if comm_world == "worker": output = self._worker_comm.all_gather(input) elif comm_world == "server": output = self._server_comm.all_gather(input) else: output = self._nodes_comm.all_gather(input) return output class RoleMakerBase(object): """ RoleMakerBase is a base class for assigning a role to current process in distributed training. A paddle developer can implement RoleMakerBase to design a role maker for worker or pserver assignment. """ def __init__(self): self._worker_endpoints = [] self._server_endpoints = [] self._role_is_generated = False self._role = None self._current_id = -1 # for heter parameter server mode self._heter_trainer_endpoints = [] self._heter_trainer_device = "CPU" self._is_heter_parameter_server_mode = False def _is_worker(self): """ return is_worker() of current process """ raise NotImplementedError("Please implement this method in child class") def _is_server(self): """ return is_server() of current process """ raise NotImplementedError("Please implement this method in child class") def _is_first_worker(self): """ Check whether the node is the first instance of worker. Returns: bool: True if this is the first node of worker, False if not. """ raise NotImplementedError("Please implement this method in child class") def _worker_num(self): """ Get current total worker number. Returns: int: worker number """ raise NotImplementedError("Please implement this method in child class") def _server_num(self): """ Get current total server number. Returns: int: server number """ raise NotImplementedError("Please implement this method in child class") def _worker_index(self): """ Get current worker id. Returns: int: node id """ raise NotImplementedError("Please implement this method in child class") def _server_index(self): """ Get current server id. Returns: int: node id """ raise NotImplementedError("Please implement this method in child class") def _role_id(self): """ Get current id. Returns: int: node id """ raise NotImplementedError("Please implement this method in child class") def _node_num(self): """ Get the training node number Returns: int: node num """ raise NotImplementedError("Please implement this method in child class") def _get_trainer_endpoints(self): """ return trainer endpoints """ return self._worker_endpoints def _get_pserver_endpoints(self): """ return pserver endpoints """ return self._server_endpoints def to_string(self): return "role: {}, current_id: {}, worker_endpoints: {}, server_endpoints: {}".format( self._role, self._current_id, self._worker_endpoints, self._server_endpoints) def _all_gather(self, input, comm_world="worker"): print("warning: RoleMakerBase does not have all gather worker.") return None def _all_reduce(self, input, mode="sum", comm_world="worker"): """ Args: input(list/numpy.array): array of one dim output(list/numpy.array): array of one dim mode(str): "sum" or "min" or "max" """ print("warning: RoleMakerBase does not have all reduce worker.") return None def _barrier(self, comm_world): """ barrier between trainers if current role is TRAINER """ print("warning: RoleMakerBase does not have barrier worker.") def _is_heter_worker(self): """ Return is_heter_worker() of current process """ warnings.warn("RoleMakerBase does not have function: _is_heter_worker.") return False def _heter_worker_num(self): """ Get current total heter-worker number. Returns: int: heter_worker number """ warnings.warn( "RoleMakerBase does not have function: _heter_worker_num.") return 0 def _get_heter_worker_endpoints(self): """ Returns: string: all heter_trainers'endpoints """ assert self._heter_trainer_endpoints != [], "Heter Worker Endpoints Not initialized" return self._heter_trainer_endpoints def _get_heter_worker_endpoint(self): """ Returns: int: corresponding heter_trainer's endpoint e.g: if we have 4 cpu-trainer(default), 2 gpu-trainer(heter) then No.0 and No.2 cpu-trainer will work with No.0 gpu-trainer and No.1 and No.3 cpu-trainer will work with No.1 gpu-trainer """ assert self._heter_trainer_endpoints != [], "Heter Worker Endpoints Not initialized" return self._heter_trainer_endpoints[(self._current_id) % self._heter_worker_num()] class PaddleCloudRoleMaker(RoleMakerBase): def __init__(self, is_collective=False, **kwargs): super(PaddleCloudRoleMaker, self).__init__() self._is_collective = is_collective self._non_distributed = False self._kwargs = kwargs self._role_is_generated = False self._server_endpoints = [] self._worker_endpoints = [] self._gloo = Gloo() # gloo instance def _barrier(self, comm_world): self._gloo.barrier(comm_world) def _all_gather(self, input, comm_world="worker"): return self._gloo.all_gather(input, comm_world) def _all_reduce(self, input, mode="sum", comm_world="worker"): return self._gloo.all_reduce(input, mode, comm_world) def _is_worker(self): """ whether current process is worker """ if not self._role_is_generated: self._generate_role() return self._role == Role.WORKER def _is_server(self): """ whether current process is server """ if not self._role_is_generated: self._generate_role() return self._role == Role.SERVER def _is_first_worker(self): """ whether current process is worker of rank 0 """ if not self._role_is_generated: self._generate_role() return self._role == Role.WORKER and self._current_id == 0 def _worker_index(self): """ get index of current worker """ if not self._role_is_generated: self._generate_role() return self._current_id def _server_index(self): """ get index of current server """ if not self._role_is_generated: self._generate_role() return self._current_id def _role_id(self): """ get index of current node """ if not self._role_is_generated: self._generate_role() return self._current_id def _worker_num(self): """ retrun the current number of worker """ if not self._role_is_generated: self._generate_role() return self._trainers_num def _server_num(self): """ return the current number of server """ if not self._role_is_generated: self._generate_role() return len(self._get_pserver_endpoints( )) if self._get_pserver_endpoints() is not None else 0 def _node_num(self): """ return the training node number """ if not self._role_is_generated: self._generate_role() return self._nodes_num def _get_trainer_endpoints(self): """ get endpoint of all trainers """ if not self._role_is_generated: self._generate_role() return self._worker_endpoints def _get_pserver_endpoints(self): """ get endpoint of all pservers """ if not self._role_is_generated: self._generate_role() return self._server_endpoints def _is_non_distributed(self): """ Return True if indispensable environment for fleetrun is not found (use python-run to launch fleet-code directly) """ if not self._role_is_generated: self._generate_role() return self._non_distributed def _heter_worker_num(self): """ get heter worker nums """ if not self._role_is_generated: self._generate_role() return self._heter_trainers_num def _is_heter_worker(self): """ whether current process is heter worker """ if not self._role_is_generated: self._generate_role() return self._role == Role.HETER_WORKER def _ps_env(self): # Environment variable PADDLE_PSERVERS_IP_PORT_LIST must be set # format: string(ip:port,ip:port), eg. 127.0.0.1:6001,127.0.0.1:6002 self._server_endpoints = os.getenv("PADDLE_PSERVERS_IP_PORT_LIST", None) if self._server_endpoints is None: # back to non_distributed execution. self._server_endpoints = "" self._trainers_num = 1 self._role = Role.WORKER self._current_id = 0 self._nodes_num = 1 self._heter_trainers_num = 0 self._heter_trainer_endpoints = None self._non_distributed = True return self._server_endpoints = self._server_endpoints.split(",") self._worker_endpoints = os.getenv("PADDLE_TRAINER_ENDPOINTS", None) if self._worker_endpoints != None: self._worker_endpoints = self._worker_endpoints.split(",") else: self._worker_endpoints = [] trainers_num = os.getenv("PADDLE_TRAINERS_NUM", None) if trainers_num == None: raise ValueError( "Can not find PADDLE_TRAINERS_NUM, please check your environment." ) trainers_num = int(trainers_num) training_role = os.getenv("TRAINING_ROLE", None) if training_role == None: raise ValueError( "Can not find TRAINING_ROLE, please check your environment.") if training_role not in ["TRAINER", "PSERVER", "HETER_TRAINER"]: raise ValueError( "TRAINING_ROLE must be PSERVER or TRAINER or HETER_TRAINER, but get {}, please check your environment.". format(training_role)) # For heter parameter server env setting heter_trainer_eplist = os.getenv("PADDLE_HETER_TRAINER_IP_PORT_LIST", "") if heter_trainer_eplist != "": try: heter_trainer_eplist = os.environ[ "PADDLE_HETER_TRAINER_IP_PORT_LIST"].split(",") except: raise ValueError( "Can not Find PADDLE_HETER_TRAINER_IP_PORT_LIST in env or its format doesn't match the requirement: 'IP:PORT,IP:PORT' ." ) self._is_heter_parameter_server_mode = True heter_trainers_num = len(heter_trainer_eplist) else: self._is_heter_parameter_server_mode = False heter_trainers_num = 0 if training_role == "TRAINER": role = Role.WORKER current_id = os.getenv("PADDLE_TRAINER_ID", None) if current_id == None: raise ValueError( "Can not find PADDLE_TRAINER_ID, please check your environment." ) current_id = int(current_id) if len(self._worker_endpoints) > 0: self._cur_endpoint = self._worker_endpoints[current_id] elif training_role == "PSERVER": role = Role.SERVER port = os.getenv("PADDLE_PORT", None) if port == None: raise ValueError( "Can not find PADDLE_PORT, please check your environment.") ip = os.getenv("POD_IP", None) if ip == None: raise ValueError( "Can not find POD_IP, please check your environment.") self._cur_endpoint = ip + ":" + port current_id = self._server_endpoints.index(self._cur_endpoint) elif training_role == "HETER_TRAINER": role = Role.HETER_WORKER cur_port = os.getenv("PADDLE_PORT", None) if cur_port == None: raise ValueError( "Can not find PADDLE_PORT, please check your environment.") cur_ip = os.getenv("POD_IP", None) if cur_ip == None: raise ValueError( "Can not find POD_IP, please check your environment.") curr_endpoint = ":".join([cur_ip, cur_port]) current_id = heter_trainer_eplist.index(curr_endpoint) self._trainers_num = trainers_num self._role = role self._current_id = current_id self._nodes_num = len( set([x.split(':')[0] for x in self._worker_endpoints])) self._heter_trainers_num = heter_trainers_num self._heter_trainer_endpoints = heter_trainer_eplist def _collective_env(self): self._current_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) self._training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") assert (self._training_role == "TRAINER") self._role = Role.WORKER self._worker_endpoints = os.getenv("PADDLE_TRAINER_ENDPOINTS") self._cur_endpoint = os.getenv("PADDLE_CURRENT_ENDPOINT") if self._worker_endpoints is None: # back to non_distributed execution. self._worker_endpoints = "127.0.0.1:6170" self._cur_endpoint = self._worker_endpoints self._non_distributed = True self._worker_endpoints = self._worker_endpoints.split(",") self._trainers_num = len(self._worker_endpoints) self._nodes_num = len( set([x.split(':')[0] for x in self._worker_endpoints])) def _gloo_init(self): # PADDLE_WITH_GLOO 1: trainer barrier, 2: all barrier use_gloo = int(os.getenv("PADDLE_WITH_GLOO", "0")) if use_gloo not in [1, 2]: return # PADDLE_GLOO_RENDEZVOUS 1: HDFS 2: FILE 3: HTTP rendezvous_type = int(os.getenv("PADDLE_GLOO_RENDEZVOUS", "0")) prefix = os.getenv("SYS_JOB_ID", "") if rendezvous_type not in [ Gloo.RENDEZVOUS.HDFS, Gloo.RENDEZVOUS.HTTP, Gloo.RENDEZVOUS.FILE ]: raise ValueError(self._gloo._err_type) need_init_all = True if use_gloo == 2 else False if rendezvous_type == Gloo.RENDEZVOUS.HDFS: dfs_name = os.getenv("PADDLE_GLOO_FS_NAME", "") dfs_ugi = os.getenv("PADDLE_GLOO_FS_UGI", "") dfs_path = os.getenv("PADDLE_GLOO_FS_PATH", "") kwargs = { "dfs.name": dfs_name, "dfs.ugi": dfs_ugi, "dfs.path": dfs_path, "store.prefix": prefix, } elif rendezvous_type == Gloo.RENDEZVOUS.HTTP: start_http_server = False manager = Manager() http_server_d = manager.dict() http_server_d["running"] = False if self._is_collective: ep_rank_0 = self._worker_endpoints[0] if self._is_first_worker(): start_http_server = True else: ep_rank_0 = os.getenv("PADDLE_GLOO_HTTP_ENDPOINT", "") if self._server_index() == 0: start_http_server = True ip, port = ep_rank_0.split(':') kwargs = { "http.host": ip, "http.port": port, "store.prefix": prefix, 'start_http_server': start_http_server, 'http_server_d': http_server_d, } else: dfs_path = os.getenv("PADDLE_GLOO_FS_PATH", "") kwargs = { "dfs.path": dfs_path, "store.prefix": prefix, } if rendezvous_type == Gloo.RENDEZVOUS.HDFS: type = "HDFS" elif rendezvous_type == Gloo.RENDEZVOUS.HTTP: type = "HTTP" else: type = "FILE" print("Gloo init with {}: need_init_all: {}, args: {}".format( type, need_init_all, kwargs)) self._gloo.init( rendezvous=rendezvous_type, role=self._role, role_id=self._role_id(), worker_num=self._worker_num(), server_num=self._server_num(), need_init_all=need_init_all, kwargs=kwargs) if rendezvous_type == Gloo.RENDEZVOUS.HTTP: http_server_d['running'] = False def _generate_role(self): """ generate role for role maker """ if not self._role_is_generated: if not self._is_collective: self._ps_env() else: self._collective_env() self._role_is_generated = True if not paddle.fluid.framework.in_dygraph_mode(): self._gloo_init() class UserDefinedRoleMaker(PaddleCloudRoleMaker): def __init__(self, is_collective=False, init_gloo=False, **kwargs): super(UserDefinedRoleMaker, self).__init__( is_collective=is_collective, init_gloo=init_gloo, **kwargs) self._init_gloo = init_gloo def _user_defined_ps_env(self): self._server_endpoints = self._kwargs.get("server_endpoints") self._worker_endpoints = self._kwargs.get("worker_endpoints", []) self._trainers_num = self._kwargs.get("worker_num", 0) if self._trainers_num == 0: assert (len(self._worker_endpoints) > 0) self._trainers_num = len(self._worker_endpoints) self._role = self._kwargs.get("role") self._current_id = self._kwargs.get("current_id") if self._role == Role.WORKER and len( self._worker_endpoints) > self._current_id: self._cur_endpoint = self._worker_endpoints[self._current_id] elif self._role == Role.SERVER: self._cur_endpoint = self._server_endpoints[self._current_id] self._nodes_num = len( set([x.split(':')[0] for x in self._worker_endpoints])) def _user_defined_collective_env(self): self._worker_endpoints = self._kwargs.get("worker_endpoints") self._current_id = self._kwargs.get("current_id") self._trainers_num = len(self._worker_endpoints) self._training_role = Role.WORKER self._nodes_num = len( set([x.split(':')[0] for x in self._worker_endpoints])) def _generate_role(self): """ generate role for role maker """ if not self._role_is_generated: if not self._is_collective: self._user_defined_ps_env() else: self._user_defined_collective_env() self._role_is_generated = True