-
Notifications
You must be signed in to change notification settings - Fork 5.5k
/
parallel_with_gloo.py
executable file
·253 lines (204 loc) · 8.05 KB
/
parallel_with_gloo.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
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
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except jin 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.
import time
from multiprocessing import Manager, Process
from paddle.distributed.fleet.base.private_helper_function import (
wait_server_ready,
)
# deprecated module import
# (TODO: GhostScreaming) It will be removed later.
from paddle.fluid import core
__all__ = []
_global_gloo_ctx = None
def _start_kv_server(port, http_server_d, size):
from paddle.distributed.fleet.utils.http_server import KVServer
http_server = KVServer(int(port), size=size)
http_server.start()
wait_seconds = 3
while http_server_d.get("running", False) or not http_server.should_stop():
time.sleep(wait_seconds)
http_server.stop()
def gloo_init_parallel_env(rank_id, rank_num, server_endpoint):
"""
Initialize parallel environment with gloo for cpu only.
Args:
- rank_id(int, required) - the index of current rank;
- rank_num (int, required) - the number of ranks in this parallel env;
- server_endpoint (str, required) - endpoint of server to init gloo context in ip:port format;
Returns:
None
Examples:
.. code-block:: python
import paddle
import multiprocessing
from contextlib import closing
import socket
port_set = set()
def find_free_port():
def _free_port():
with closing(socket.socket(socket.AF_INET,
socket.SOCK_STREAM)) as s:
s.bind(('', 0))
return s.getsockname()[1]
while True:
port = _free_port()
if port not in port_set:
port_set.add(port)
return port
def test_gloo_init(id, rank_num, server_endpoint):
paddle.distributed.gloo_init_parallel_env(
id, rank_num, server_endpoint)
def test_gloo_init_with_multiprocess(num_of_ranks):
jobs = []
server_endpoint = "127.0.0.1:%s" % (find_free_port())
for id in range(num_of_ranks):
p = multiprocessing.Process(
target=test_gloo_init,
args=(id, num_of_ranks, server_endpoint))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
if __name__ == '__main__':
# Arg: number of ranks (processes)
test_gloo_init_with_multiprocess(2)
"""
assert (
rank_num < 2
) is False, "rank_num should greater than or equal to 2 for parallel environment initialzation."
# init gloo context
manager = Manager()
# global dict to store status
http_server_status = manager.dict()
http_server_status["running"] = False
if rank_id == 0:
# The scope for worker used by http server is '_worker'
size = {'_worker': rank_num}
http_server_proc = Process(
target=_start_kv_server,
args=(int(server_endpoint.split(":")[1]), http_server_status, size),
)
http_server_proc.daemon = True
http_server_status["running"] = True
http_server_proc.start()
# all processes in this parallel environment should wait until server is ready
wait_server_ready([server_endpoint])
gloo_strategy = core.GlooParallelStrategy()
gloo_strategy.rank = rank_id
gloo_strategy.rank_num = rank_num
gloo_strategy.ip_address = server_endpoint.split(":")[0]
gloo_strategy.ip_port = int(server_endpoint.split(":")[1])
# default_init_timeout_seconds
gloo_strategy.init_seconds = 3600
# default_run_timeout_seconds
gloo_strategy.run_seconds = 9999999
global _global_gloo_ctx
_global_gloo_ctx = core.GlooParallelContext(gloo_strategy)
_global_gloo_ctx.init()
if rank_id == 0:
http_server_status["running"] = False
http_server_proc.join()
def gloo_barrier():
"""
Call barrier function with initialized gloo context.
Args:
None
Returns:
None
Examples:
.. code-block:: python
import paddle
import multiprocessing
from contextlib import closing
import socket
port_set = set()
def find_free_port():
def _free_port():
with closing(socket.socket(socket.AF_INET,
socket.SOCK_STREAM)) as s:
s.bind(('', 0))
return s.getsockname()[1]
while True:
port = _free_port()
if port not in port_set:
port_set.add(port)
return port
def test_gloo_barrier(id, rank_num, server_endpoint):
paddle.distributed.gloo_init_parallel_env(
id, rank_num, server_endpoint)
paddle.distributed.gloo_barrier()
def test_gloo_barrier_with_multiprocess(num_of_ranks):
jobs = []
server_endpoint = "127.0.0.1:%s" % (find_free_port())
for id in range(num_of_ranks):
p = multiprocessing.Process(
target=test_gloo_barrier,
args=(id, num_of_ranks, server_endpoint))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
if __name__ == '__main__':
# Arg: number of ranks (processes)
test_gloo_barrier_with_multiprocess(2)
"""
assert _global_gloo_ctx is not None, "gloo context is not initialzed."
_global_gloo_ctx.barrier()
def gloo_release():
"""
Release the parallel environment initialized by gloo
Args:
None
Returns:
None
Examples:
.. code-block:: python
import paddle
import multiprocessing
from contextlib import closing
import socket
port_set = set()
def find_free_port():
def _free_port():
with closing(socket.socket(socket.AF_INET,
socket.SOCK_STREAM)) as s:
s.bind(('', 0))
return s.getsockname()[1]
while True:
port = _free_port()
if port not in port_set:
port_set.add(port)
return port
def test_gloo_release(id, rank_num, server_endpoint):
paddle.distributed.gloo_init_parallel_env(
id, rank_num, server_endpoint)
paddle.distributed.gloo_barrier()
paddle.distributed.gloo_release()
def test_gloo_release_with_multiprocess(num_of_ranks):
jobs = []
server_endpoint = "127.0.0.1:%s" % (find_free_port())
for id in range(num_of_ranks):
p = multiprocessing.Process(
target=test_gloo_release,
args=(id, num_of_ranks, server_endpoint))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
if __name__ == '__main__':
# Arg: number of ranks (processes)
test_gloo_release_with_multiprocess(2)
"""
if _global_gloo_ctx is not None:
_global_gloo_ctx.release()