collective.py 5.5 KB
Newer Older
K
kuizhiqing 已提交
1
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
2
#
K
kuizhiqing 已提交
3 4 5
# 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
6
#
K
kuizhiqing 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
8
#
K
kuizhiqing 已提交
9 10 11 12 13 14
# 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.

K
kuizhiqing 已提交
15
from .controller import Controller, ControleMode
K
kuizhiqing 已提交
16 17 18 19 20 21 22 23

import json
import os
import six
import time


class CollectiveController(Controller):
24

K
kuizhiqing 已提交
25 26
    @classmethod
    def enable(cls, ctx):
K
kuizhiqing 已提交
27
        # collective is the default mode
K
kuizhiqing 已提交
28 29
        if ctx:
            ctx.logger.debug("{} enabled".format(cls.__name__))
K
kuizhiqing 已提交
30
            ctx.args.run_mode = ControleMode.COLLECTIVE
K
kuizhiqing 已提交
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
            return True
        else:
            return False

    def build_pod(self):
        self.pod.replicas = self.pod_replicas()

        # rank will be reset when restart
        self.pod.rank = self.ctx.args.rank

        port = self.ctx.node.get_free_port()

        # compatible
        endpoints = [
            "{}:{}".format(self.ctx.node.ip, p)
            for p in self.ctx.node.get_free_ports(self.pod.replicas)
        ]

        data = json.dumps({
            'name': self.pod.name,
            'rank': self.pod.rank,
            'replicas': self.pod.replicas,
            'dtype': self.ctx.node.device.dtype,
            'candidate': '{}:{}'.format(self.ctx.node.ip, port),
            'endpoints': ",".join(endpoints),
        })

58 59 60 61
        peer_list, rank = self.master.sync_peers('/{}/info'.format(self.job.id),
                                                 self.pod.name, data,
                                                 self.job.replicas,
                                                 self.pod.rank)
K
kuizhiqing 已提交
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
        self.pod.rank = rank

        if len(peer_list) < 1:
            return False

        peer_list = [json.loads(i) for i in peer_list]

        self.ctx.logger.debug("sync peers done {}".format(peer_list))
        self.save_pod_log(peer_list)

        global_size = sum([i['replicas'] for i in peer_list])
        rank_offset = sum([i['replicas'] for i in peer_list[:rank]])
        '''
        The new designed collective need nothing but a master endpoint
        '''
        collective_master = peer_list[0]['candidate']

        job_endpoints = [i['endpoints'] for i in peer_list]

        self.pod.reset()
K
kuizhiqing 已提交
82 83
        selected_dev_key = self.ctx.node.device.get_selected_device_key()
        selected_dev_list = self.ctx.node.device.get_selected_devices(
84
            self.ctx.args.devices)
K
kuizhiqing 已提交
85 86 87 88 89 90 91
        for i in range(self.pod.replicas):
            e = {
                "PADDLE_MASTER": collective_master,
                "PADDLE_GLOBAL_SIZE": "{}".format(global_size),
                "PADDLE_LOCAL_SIZE": "{}".format(self.pod.replicas),
                "PADDLE_GLOBAL_RANK": "{}".format(i + rank_offset),
                "PADDLE_LOCAL_RANK": "{}".format(i),
K
kuizhiqing 已提交
92
                "PADDLE_NNODES": "{}".format(self.job.replicas),
K
kuizhiqing 已提交
93 94 95 96 97 98 99
                ## compatible env
                "PADDLE_TRAINER_ENDPOINTS": ",".join(job_endpoints),
                "PADDLE_CURRENT_ENDPOINT": endpoints[i],
                "PADDLE_TRAINER_ID": "{}".format(i + rank_offset),
                "PADDLE_TRAINERS_NUM": "{}".format(global_size),
                "PADDLE_RANK_IN_NODE": str(i),
            }
100 101 102 103 104
            if len(selected_dev_list) > 0:
                if self.pod.replicas == 1:
                    e.update({selected_dev_key: ",".join(selected_dev_list)})
                else:
                    e.update({selected_dev_key: selected_dev_list[i]})
105
            else:
106 107
                e.update({'PADDLE_DISTRI_BACKEND': 'gloo'})

K
kuizhiqing 已提交
108 109 110 111 112 113
            self.add_container(envs=e, log_tag=i)

        return True


class CollectiveElasticController(CollectiveController):
114

K
kuizhiqing 已提交
115 116 117 118
    @classmethod
    def enable(cls, ctx):
        if ctx.args.master and ctx.args.master.startswith("etcd://"):
            ctx.logger.debug("{} enabled".format(cls.__name__))
K
kuizhiqing 已提交
119
            ctx.args.run_mode = ControleMode.COLLECTIVE
K
kuizhiqing 已提交
120 121 122 123 124 125 126
            return True
        else:
            return False

    def register(self):
        if self.job.id == 'default':
            self.ctx.logger.warning(
127 128
                'Using default job name may cause conflict, add --job_id in args'
            )
K
kuizhiqing 已提交
129 130 131 132 133 134 135 136 137 138 139 140

        self.master.register_heartbeat(self.job.id, self.pod.name)

    def run(self):

        timeout = self.ctx.args.elastic_timeout if self.job.elastic else self.ctx.args.elastic_timeout * 10
        self.register()

        while self.pod.restart <= self.ctx.args.max_restart:

            self.build_job()

141 142
            self.ctx.logger.info("Waiting peer ready...")

143 144 145
            ok, replicas = self.master.wait_peer_ready(self.job.replicas_min,
                                                       self.job.replicas_max,
                                                       timeout)
K
kuizhiqing 已提交
146 147 148
            if ok:
                self.job.replicas = replicas
            else:
K
kuizhiqing 已提交
149
                self.ctx.logger.warning("peer not ready {}".format(self.job))
K
kuizhiqing 已提交
150 151 152 153 154 155 156 157 158
                break

            self.ctx.logger.debug("Run {}".format(self.job))

            if not self.build_pod():
                continue

            self.master.set_status(self.ctx.status.RUNNING)

159
            self.deploy_pod()
K
kuizhiqing 已提交
160 161 162 163 164

            if self.watch():
                break

        self.ctx.logger.debug("Job done {}".format(self.job))