-
Notifications
You must be signed in to change notification settings - Fork 1
/
dask_tensorflow.py
590 lines (488 loc) · 25.1 KB
/
dask_tensorflow.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
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
#!/usr/bin/env python3
#coding:utf-8
"""
@Author: Abael He<abaelhe@icloud.com>
@file: dask_tensorflow.py
@time: 7/26/19 12:05 PM
@license: All Rights Reserved, Abael.com
"""
import imp
import six, os, sys, time, collections, socket, json, datetime, traceback, signal, zipimport, threading, tornado
from functools import partial,wraps,lru_cache
from concurrent.futures import ThreadPoolExecutor
from tornado import gen, locks
from tornado import process
from tornado.gen import coroutine, Return, Future
from tornado.locks import Event, Condition, Lock, Semaphore, BoundedSemaphore
from distributed import (Client, Worker, get_worker, secede as worker_secede, get_client, Reschedule, Pub, Sub, Lock,
Variable, worker_client)
from distributed.protocol import to_serialize
from distributed.threadpoolexecutor import secede as thread_pool_secede
from distributed.worker import thread_state
import logging
logger = logging.getLogger('distributed.preloading')
from dask_usage import USAGE_INFO
from dask_signal import IN_DASK, GLOBAL_IOLOOP
from dask_global import (cuda_free_indexes, global_cluster, urlparse, safe_cmd, model_cleanup, async_send,
DASK_PYTHONHASHSEED, DASK_PYTHON_INTERPRETER,
DASK_READ_CHUNK_SIZE, DASK_WORKSPACE, DASK_DATA_POOL_DIR, DASK_MODEL_POOL_DIR)
TF_PORT = 11111
TENSORFLOW_KEYS = ['chief', 'master', 'service', 'session', 'ps', 'worker']
TF_CONFIG_ENV = 'TF_CONFIG'
TF_TASK_ENV_KEY = 'task'
TF_TASK_TYPE_KEY = 'type'
TF_TASK_ID_KEY = 'index'
TF_CLUSTER_KEY = 'cluster'
TF_SERVICE_KEY = 'service'
TF_SESSION_MASTER_KEY = 'session_master'
TF_EVAL_SESSION_MASTER_KEY = 'eval_session_master'
TF_MODEL_DIR_KEY = 'model_dir'
TF_GRPC_SCHEME = 'grpc://'
# Tensor TF_CONFIG environment:
# Master | Chief: require 'cluster' key in TF_CONFIG;
# Worker: 'cluster' key MUST NOT in TF_CONFIG
TF_CONFIG = {
TF_TASK_ENV_KEY: {TF_TASK_TYPE_KEY: 'chief', TF_TASK_ID_KEY: 0},
TF_CLUSTER_KEY: {
'evaluator': ['192.168.1.1:2224'], # ONLY ONE `evaluator` !
'chief': ['192.168.1.1:2222'], 'master': ['192.168.1.1:2223'], # This TWO is for Super ONE
'ps': ['192.168.1.104:2222', '192.168.1.105:2222'],
'worker': ['192.168.1.100:2222', '192.168.1.101:2222',
'192.168.1.102:2222', '192.168.1.103:2222'],
},
TF_SERVICE_KEY: {},
TF_SESSION_MASTER_KEY: {},
TF_EVAL_SESSION_MASTER_KEY: {},
}
class TFTaskType(object):
MASTER = 'master'
PS = 'ps'
WORKER = 'worker'
CHIEF = 'chief'
EVALUATOR = 'evaluator' # 'For distributed training, there can only be one `evaluator` task, means: `task_id=0`'
dic = {'chief':0, 'master':1, 'ps':2, 'worker':4, 'evaluator':8}
def __init__(self, task_type): # TaskType And Device Filters association rules
self.task_type = task_type
if self._task_type == TFTaskType.MASTER:
device_filters = ['/job:ps', '/job:master']
elif self._task_type == TFTaskType.CHIEF:
device_filters = ['/job:ps', '/job:chief']
elif self._task_type == TFTaskType.WORKER:
device_filters = ['/job:ps', '/job:worker/task:%d' % self._task_id]
elif self._task_type == TFTaskType.PS:
device_filters = ['/job:ps', '/job:worker', '/job:chief', '/job:master']
else:
# If the task_type is `EVALUATOR` or something other than the ones in
# TaskType then don't set any device filters.
device_filters = None
self.device_filters = device_filters
# ref.: distributed:worker.py:run(server, comm, function, args=(), kwargs={}, is_coro=None, wait=True):
# support builtin special kwargs: dask_worker:server, dask_scheduler:server, ONLY available: distributed.client._run()
#################################
def gpu_filter(_devices, gpu_flag):
return [x for x in _devices if x['name'].find('GPU') > 0] if gpu_flag is True else _devices
def tensorflow_devices():
from tensorflow.python.client import device_lib
device_list = device_lib.list_local_devices()
for device in device_list:
locality = device.locality
yield (device.name, {
'name': device.name,
'memory_limit': device.memory_limit,
'device_type': device.device_type,
'byte_size': device.ByteSize(),
'physical_device_desc': device.physical_device_desc,
'incarnation': device.incarnation,
'locality': {
'bus_id': locality.bus_id,
'numa_node': locality.numa_node,
'links': str(locality.links)
},
})
def tensorflow_options(gpu_mem_fraction=0.95):
import tensorflow as tf
Config = tf.compat.v1.ConfigProto
print('TENSORFLOW JIT STATUS: %s' % tf.config.optimizer.get_jit())
tf.config.optimizer.set_jit(1)
tf_option = tf.compat.v1.ConfigProto(log_device_placement=True, allow_soft_placement=True,
gpu_options=tf.compat.v1.GPUOptions(
per_process_gpu_memory_fraction=gpu_mem_fraction,
force_gpu_compatible=True,
allow_growth=True,
),
)
tf_option.graph_options.optimizer_options.opt_level = tf.compat.v1.OptimizerOptions.L1
tf_option.graph_options.optimizer_options.global_jit_level = tf.compat.v1.OptimizerOptions.ON_1
tf_option.graph_options.optimizer_options.do_function_inlining = True
return tf_option
#s=tf_option.SerializeToString();tf.compat.v1.ConfigProto.FromString(s)
class TFActor(object):
def __del__(self):
self.recovery()
def __str__(self):
return "<%s %s>" %(self.__class__.__name__, self.key)
def __init__(self, key, *args, tf_config=None, tf_option=None, scheduler_info=None, **kwargs):
# here we made this thread an OWNER of this task by secede from it's ThreadPoolExecutor.
# NOTE: `thread_pool_secede` ONLY works in NON-coroutine actor_exectutor, ref:worker.actor_execute()
self.dask_worker = get_worker()
self.thrid = threading.get_ident()
thread_pool_secede(adjust=True)
self.dask_worker.loop.add_callback(self.dask_worker.transition, thread_state.key, "long-running")
self.key = key
self.name = self.dask_worker.name
self.hostname = socket.gethostname()
self.address = self.dask_worker.address
self.scheduler_info = scheduler_info
model_name = self.key.partition(':')[0]
self.model_name = model_name[:-4] if model_name.endswith('.zip') else model_name
self.tf_option = json.loads(tf_option) if isinstance(tf_option, str) else tf_option
self.tf_config = json.loads(tf_config) if isinstance(tf_config, str) else tf_config
self.dask_cwd = os.path.abspath(os.getcwd())
self.tf_model_pool_dir = os.path.abspath(DASK_MODEL_POOL_DIR)
self.tf_data_pool_dir = os.path.abspath(DASK_DATA_POOL_DIR)
self.tf_data_dir = os.path.join(self.tf_data_pool_dir, self.model_name)
self.tf_config_dir = os.path.join(self.tf_data_dir, 'config')
self.tf_save_dir = os.path.join(self.tf_data_dir, 'ckpt')
self.tf_log_dir = os.path.join(self.tf_data_dir, 'log')
os.system('mkdir -p %r; rm -rf %r; mkdir -p %r %r %r %r' % (self.tf_save_dir, self.tf_save_dir,
self.tf_data_dir, self.tf_config_dir, self.tf_save_dir, self.tf_log_dir))
os.chdir(self.tf_data_dir)
self.sys_stdio = (sys.__stdin__, sys.__stdout__, sys.__stderr__, sys.stdin, sys.stdout, sys.stderr)
self.stdout = open(os.path.join(self.tf_log_dir, '%s.log' % self.key.partition(':')[-1]), 'a+', encoding=sys.stdout.encoding)
sys.__stdout__ = sys.__stderr__ = sys.stdout = sys.stderr = self.stdout
self.stdin = sys.stdin
self.sys_path, self.sys_argv = sys.path[:], sys.argv[:]
logger.info('Accepted Tensorflow Key:%s, Job:%s, Options:%s, Scheduler:%s', key, tf_config, tf_option, scheduler_info)
self.devices = dict(tensorflow_devices())
self.future_chunk_size = DASK_READ_CHUNK_SIZE
self.args = args
self.kwargs = kwargs
self.sub = Sub(self.key, worker=self.dask_worker)
self.result = Pub(model_name, worker=self.dask_worker)
self.exe = self.preflight()
self.dask_worker.loop.add_callback(self.flight, self.exe)
def device_info(self, xla=None, gpu=True):
if xla is None:
return gpu_filter([v for (x, v) in self.devices.items() if v['name'].find('GPU') >= 0], gpu_flag=gpu)
elif xla is True:
return gpu_filter([v for (x, v) in self.devices.items() if v['name'].find('XLA') >= 0], gpu_flag=gpu)
else:
return gpu_filter([v for (x, v) in self.devices.items() if v['name'].find('XLA') < 0], gpu_flag=gpu)
def tensorflow_env(self, tf_option, tf_config, dask_context, cuda_indexes=None):
model_entrypoint = os.path.join(self.tf_model_pool_dir, self.model_name)
zip_ep, pkg_ep =model_entrypoint + '.zip', os.path.join(model_entrypoint, '__main__.py')
if os.path.exists(pkg_ep) and os.path.isfile(pkg_ep):
model_entrypoint = pkg_ep
elif os.path.exists(zip_ep) and os.path.isfile(zip_ep):
model_entrypoint = zip_ep
else:
raise Exception(USAGE_INFO)
env_dict = {}
for key in ('LANG', 'PATH', 'CUDA_HOME', 'LD_LIBRARY_PATH',
'USER', 'HOME', 'HOSTNAME', 'SHELL', 'TERM', 'SHLVL', 'MAIL', 'SSH_CONNECTION', 'SSH_TTY', 'SSH_CLIENT'):
val = os.getenv(key)
if val is not None:
env_dict[key] = val
env_dict.update(
XLA_FLAGS='--xla_hlo_profile',
TF_DASK_PID=str(os.getpid()),
RF_DASK_PGRP=str(os.getpgrp()),
TF_XLA_FLAGS=("--tf_xla_cpu_global_jit " + os.environ.get("TF_XLA_FLAGS", "")),
TF_MODEL=self.model_name,
TF_CONTEXT=json.dumps(dask_context),
TF_CONFIG=json.dumps(tf_config),
TF_MODEL_POOL_DIR=self.tf_model_pool_dir,
TF_DATA_POOL_DIR=self.tf_data_pool_dir,
TF_MODEL_ENTRYPOINT=model_entrypoint,
TF_CONFIG_DIR=self.tf_config_dir,
TF_DATA_DIR=self.tf_data_dir,
TF_LOG_DIR=self.tf_log_dir,
TF_SAVE_DIR=self.tf_save_dir,
PYTHONPATH=':'.join([self.tf_model_pool_dir, self.tf_data_dir, self.dask_cwd]),
PYTHONHASHSEED=str(int(DASK_PYTHONHASHSEED)),
PYTHONIOENCODING=sys.getdefaultencoding(),
PYTHONUNBUFFERED='True',
)
if cuda_indexes: # we explicitly assign GPU indexes to use; let tensorflow aware of ONLY these indexes
env_dict['CUDA_VISIBLE_DEVICES'] = cuda_indexes
return env_dict
def log(self, msg, *args, flush=True):
self.stdout.write((msg % args) if args else msg)
if flush:
self.stdout.flush()
def run_model(self, stdin, stdout, *args, **kwargs):
import sys
sys.stdin = stdin
self.stdout = sys.stdout = sys.stderr = sys.__stdout__ = sys.__stderr__ = stdout
self.log('HERE IN ASYNC SUBPROCESS: %s' % os.getpid())
model_name = os.getenv('TF_MODEL')
model_entry =os.getenv('TF_MODEL_ENTRYPOINT')
if model_entry.endswith('.zip'):
model_root, modname = model_entry, '__main__'
elif model_entry.endswith('.py'):
model_root, modname = os.path.dirname(model_entry), os.path.basename(model_entry).rsplit('.',1)[0]
self.log('HERE IN ASYNC MODEL START, %s, %s' % (modname, model_root))
sys.path.insert(0, model_root)
__import__(modname)
def preflight(self):
# this NODE is selected for this task
node_name, node_port, cuda_indexes, dask_url = self.tf_config.pop('dask').split(':', 3)
job_name, task_index = self.tf_config['task']['type'], self.tf_config['task']['index']
tensorflow_addr = self.tf_config['cluster'][job_name][task_index]
using_xla_gpu_devices = self.device_info(xla=True, gpu=True)
using_xla_gpu_device_names = sorted([x['name'] for x in using_xla_gpu_devices])
if isinstance(self.tf_option, (str, bytes)):
import tensorflow as tf
tf_option = tf.compat.v1.ConfigProto.FromString(self.tf_option)
elif self.tf_option is not None:
tf_option = self.tf_option
else:
tf_option = tensorflow_options()
dask_context = {
'model_task': '%s, %s' % (self.key, ','.join(using_xla_gpu_device_names)),
'model_addr': tensorflow_addr,
'worker_addr': self.address,
'schduler_addr': self.scheduler_info,
'workspace': DASK_WORKSPACE,
'local_dir': self.dask_cwd,
'pid': os.getpid(),
'thread_id': self.thrid,
'code': 0,
}
env_dict = self.tensorflow_env(tf_option, self.tf_config, dask_context, cuda_indexes=cuda_indexes)
cmd = [sys.executable, r'-u', env_dict['TF_MODEL_ENTRYPOINT'], self.key]
fmt = 'Model Start, key:%s,\n cmd:%s\n dask_context:%s\n sys.path:%s\n tf_option:%s\n tf_config:%s\n\n'
self.log(fmt % (self.key, cmd, dask_context, self.sys_path, tf_option, self.tf_config))
for k, v in env_dict.items():
if not isinstance(k, str) or not isinstance(v, str):
self.log('Error env k:%s, v:%s\n' % (k, v))
exe_package = partial(process.Subprocess, cmd, executable=DASK_PYTHON_INTERPRETER,
cwd=env_dict['TF_DATA_DIR'], env=env_dict, preexec_fn=None,
stdin=self.stdin, stdout=self.stdout, stderr=self.stdout, encoding=sys.getdefaultencoding(),
pass_fds=(self.stdin.fileno(), self.stdout.fileno()), universal_newlines=False, bufsize=0,
restore_signals=False, start_new_session=False)
return exe_package
def flight(self, exe_package):
# flighting in main thread, since `SIGCHLD` MUST received in it; and then correctly call exit callback.
self.exe = exe_package()
self.exe.set_exit_callback(self.landed)
msg = '\n ***** Tensorflow Task Inited, key:%s, sub:%s, pid:%s ***** ' %(self.key, self.exe.pid, os.getpid())
self.log(msg)
def landed(self, retval=0):
self.log("worker pub msg: %s", {self.key: retval})
self.result.put({self.key: retval})
ident = yield self.dask_worker.scheduler.identity()
msg = yield self.sub._get(timeout=10)
self.log('Tensorflow Push Message Received, sub:%s, msg:%s, ident:%s' % (self.key, msg, ident))
msg = '\n ***** Tensorflow Task Finished, key:%s, ret:%s, tid:%s, pid:%s, ***** \n\n' % (
self.key, retval, threading.get_ident(), os.getpid())
self.log(msg)
self.recovery()
def recovery(self):
if self.sys_stdio is None:
return
self.log('State Recovery:%s', self.key)
os.chdir(self.dask_cwd)
sys.__stdin__, sys.__stdout__, sys.__stderr__, sys.stdin, sys.stdout, sys.stderr = self.sys_stdio
sys.path, sys.argv = self.sys_path, self.sys_argv
if self.stdin:
if self.stdin != sys.__stdin__:
self.stdin.close()
else:
self.stdin.flush()
if self.stdout:
if self.stdout != sys.__stdout__:
self.stdout.close()
else:
self.stdout.flush()
self.stdin = self.stdout = self.sys_stdio = self.sys_path = self.sys_argv = self.dask_worker = None
del self.result
del self.exe
del self.sub
def tensorflow_gen_config(free_node_name_map=None, parallel=1, save='~/tf_configs.json', **tf_cluster_spec):
if free_node_name_map:
r = [(url, name, indexes) for url, (name, indexes) in free_node_name_map.items() if len(indexes) > 0]
else:
client = global_cluster(asynchronous=False)
r =[(url, name, indexes) for url, (name, indexes) in client.run(cuda_free_indexes).items() if len(indexes) > 0]
free_node_urls = sorted(r, key=lambda x: -len(x[-1]))
free_gpu_devices = collections.deque([(url, name, x) for (url, name, indexes) in free_node_urls for x in indexes])
total_gpu_cores = len(free_gpu_devices)
if parallel > 1:
total_gpu_cores = max(1, int(total_gpu_cores // parallel))
# tf_cluster_spec['worker'] = max(1, (len(free_gpu_devices)-2) // 10)# 预计同时会有10个同事跑GPU任务; 公平调度策略, 预留10个;
if total_gpu_cores < 1:
raise Exception('All Machines is busy, Total Available:%s' % len(free_node_urls))
if len(tf_cluster_spec) < 1:
if total_gpu_cores == 1:
tf_cluster_spec ={'chief': 1}
elif total_gpu_cores == 2:
tf_cluster_spec ={'chief': 1, 'worker':1}
elif total_gpu_cores == 3:
tf_cluster_spec ={'chief': 1, 'worker':2}
else:
tf_cluster_spec ={'chief': 1, 'ps': 1, 'worker': max(1, len(free_gpu_devices)-2)}
print('tf_cluster_spec: %s' % tf_cluster_spec)
tf_cluster = collections.defaultdict(list)
port_allocation = collections.defaultdict(lambda : (TF_PORT - 1))
chief_allocation = tf_cluster_spec.pop('chief', 1)
ps_allocation = tf_cluster_spec.pop('ps', 0)
tf_configs = []
if chief_allocation:
job_name = 'chief'
job_task_index = 0
chief_node_url, chief_node_name, cuda_index = free_gpu_devices.pop()
chief_node_host, _ = urlparse(chief_node_url).netloc.rsplit(':', 1)
port_allocation[chief_node_host] += 1
chief_node_port = port_allocation[chief_node_host]
tf_cluster[job_name].append('%s:%s' % (chief_node_host, chief_node_port))
tf_configs.append({
'cluster': tf_cluster,
'task': {'type': job_name, 'index': job_task_index},
'dask':'%s:%s:%s:%s' %(chief_node_name, chief_node_port, cuda_index, chief_node_url)
})
if ps_allocation:
job_name = 'ps'
job_task_index = 0
ps_node_url, ps_node_name, cuda_index = free_gpu_devices.pop()
ps_node_host, _ = urlparse(ps_node_url).netloc.rsplit(':', 1)
port_allocation[ps_node_host] += 1
ps_node_port = port_allocation[ps_node_host]
tf_cluster[job_name].append('%s:%s' % (ps_node_host, ps_node_port))
tf_configs.append({
'cluster': tf_cluster,
'task': {'type': job_name, 'index': job_task_index},
'dask': '%s:%s:%s:%s' % (ps_node_name, ps_node_port, cuda_index, ps_node_url)
})
for job_name, machine_total in tf_cluster_spec.items():
for job_task_index in range(machine_total):
dask_url, node_name, cuda_index = free_gpu_devices.popleft()
dask_host, _ = urlparse(dask_url).netloc.rsplit(':', 1)
port_allocation[dask_host] += 1
node_port = port_allocation[dask_host]
tf_cluster[job_name].append('%s:%s' % (dask_host, node_port))
tf_configs.append({
'cluster': tf_cluster,
'task': {'type': job_name, 'index': job_task_index},
'dask': '%s:%s:%s:%s' % (node_name, node_port, cuda_index, dask_url)
})
if save:
with open(os.path.expanduser(save), 'wb') as jsf:
jsf.write(json.dumps(tf_configs).encode())
jsf.flush()
return tf_configs
def dask_sork_key(task_key):
m, _, n = task_key.partition('@')
(model_name, job_name, job_index), (node_name, cuda_index) = m.split(':'), n.split(':')
return model_name, TFTaskType.dic[job_name], int(job_index), node_name, cuda_index
@coroutine
def startup_actors(scheduler_info, client, model_name, tf_option, tf_configs, future):
rsc = {'CUDA_GPU': 1}
def gen_func(tf_config):
job_name, job_index = tf_config['task']['type'], tf_config['task']['index']
node_name, task_port, cuda_index, node_url = tf_config['dask'].split(':', 3)
task_key = '%s:%s:%s@%s:%s' % (model_name, job_name, job_index, node_name, cuda_index)
actor_startup = partial(client.submit, TFActor, task_key,
tf_config =tf_config, tf_option =tf_option, scheduler_info=scheduler_info,
key=task_key, workers=[node_url], resources=rsc,
fifo_timeout="100 ms",
retries=10,
priority=100,
allow_other_workers=False,
actor=True)
return task_key, actor_startup
startups =[gen_func(tf_config) for tf_config in tf_configs]
logger.info('Submitting: %s, Resources:%s',[x[0] for x in startups], rsc)
actors = yield {k:s() for (k, s) in startups}
future.set_result(actors)
@coroutine # eg.: job_counts={'ps':10, 'workers':100}, ParameterServers:10, CUDAworkers:100
def tensorflow_scheduler(global_future, model_name, client=None, tf_option=None, tf_port=None, **tf_cluster_spec):
scheduler_info =yield client.scheduler.identity()
cuda_free_map =yield client.run(cuda_free_indexes)
tf_configs = tensorflow_gen_config(free_node_name_map=cuda_free_map, **tf_cluster_spec)
logger.info('Model Schedule %s: \n tf_configs:%s\n\n', model_name, tf_configs)
tf_option = tf_option if isinstance(tf_option, (str, bytes)) else (tf_option.SerializeToString() if tf_option else tf_option)
chief_configs, ps_configs, other_configs = [], [], []
for tf_config in tf_configs:
task_type = tf_config['task']['type']
task_index = tf_config['task']['index']
if task_type in ('chief', 'master'):
chief_configs.append(tf_config)
elif task_type in ('ps',):
ps_configs.append(tf_config)
else:
other_configs.append(tf_config)
s_time = time.time()
dt = datetime.datetime.now()
chief_configs.sort(key=lambda cfg: cfg['task']['index'])
ps_configs.sort(key=lambda cfg: cfg['task']['index'])
other_configs.sort(key=lambda cfg: cfg['task']['index'])
client.loop.set_default_executor(ThreadPoolExecutor(max_workers=len(tf_configs)))
result_future = Future()
result_future.tf_configs = tf_configs
result_future.tf_option = tf_option
result_future.cuda_map = cuda_free_map
chief_future = Future()
client.loop.add_callback(startup_actors, scheduler_info, client, model_name, tf_option, tf_configs, chief_future)
chief_actors = yield chief_future
sorted_task_keys = list(sorted(chief_actors.keys(), key=lambda x:dask_sork_key(x)))
sub = Sub(model_name, client=client)
pubs ={k: Pub(model_name, client=client) for k in sorted_task_keys}
scheduler_info =yield client.scheduler.identity() # data flush sync between this client and scheduler
def chief_finish(task_key, actor, fu):
value = fu.result()
logger.info('Tensorflow Finished[%s/%s], key:%s, val:%s', len(chief_actors), len(tf_configs), task_key, value)
chief_actors[task_key] = actor
if len(chief_actors) == len(tf_configs):
logger.info('Tensorflow Cluster All Finished: %s', chief_actors.keys())
# Chief First.
msgs = {}
chief_key_actor = sorted_task_keys[0]
while (len(msgs) + 1) < len(chief_actors):
msg = yield sub._get()
logger.info('Sub Rcv %s:%s', type(msg), msg)
msgs.update(msg)
import pdb;pdb.set_trace()
# A = yield chief_actor.get_result()
assert chief_key_actor in msgs, 'Tensorflow Chief Task Required: %s' % chief_key_actor
time.sleep(1)
future = yield model_cleanup(client, model_name)
import pdb;pdb.set_trace()
logger.info("Tensorflow Task clean, %s", chief_actors)
global_future.set_result(chief_actors)
def start_tensorflow(model_name, client=None, options=None, port=TF_PORT, **kwargs):
""" Start Tensorflow on Dask Cluster
This launches Tensorflow Servers alongside Dask workers in-process
Examples
--------
>>> client = Client('dask-scheduler-address:8786')
>>> tf_spec, dask_spec = start_tensorflow(client)
>>> tf_spec.as_dict()
{'worker': ['192.168.1.100:2222', '192.168.1.101:2222']}
Specify desired number of jobs types as keyword args
>>> tf_config = tf.compat.v1.OptimizerOptions()
>>> tf_config.GlobalJitLevel = tf_config.OFF
>>> tf_config.do_function_inlining = True
>> tf_config.opt_level
>>> tf_config.gpu_options.force_gpu_compatible = True
>>> tf_spec, dask_spec = start_tensorflow(client, tf_config=tf_config, chief=1, master=1, ps=2, worker=30)
>>> tf_spec.as_dict()
{
'chief': ['192.168.1.1:2222'],
'master': ['192.168.1.1:2223'],
'ps': ['192.168.1.104:2222', '192.168.1.105:2222'],
'worker': ['192.168.1.100:2222', '192.168.1.101:2222',
'192.168.1.102:2222', '192.168.1.103:2222']
}
"""
client = client if client is not None else global_cluster(asynchronous=True)
global_future = Future()
tensorflow_scheduler_wrap = partial(tensorflow_scheduler, global_future, model_name, client=client, tf_port=port, **kwargs)
if client.asynchronous:
global_future.add_done_callback(lambda fu: client.loop.stop())
client.loop.add_callback(tensorflow_scheduler_wrap)
client.loop.start()
result = global_future.result()
else:
result = client.sync(client.loop, tensorflow_scheduler_wrap)
return result