-
Notifications
You must be signed in to change notification settings - Fork 15
/
pytorch_ec2.py
975 lines (866 loc) · 44.2 KB
/
pytorch_ec2.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
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
from __future__ import print_function
import sys
import threading
import Queue
import paramiko as pm
import boto3
import time
import json
import os
class Cfg(dict):
def __getitem__(self, item):
item = dict.__getitem__(self, item)
if type(item) == type([]):
return [x % self if type(x) == type("") else x for x in item]
if type(item) == type(""):
return item % self
return item
cfg = Cfg({
"name" : "Timeout", # Unique name for this specific configuration
"key_name": "HongyiScript", # Necessary to ssh into created instances
# Cluster topology
"n_masters" : 1, # Should always be 1
"n_workers" : 8,
"num_replicas_to_aggregate" : "8",
"method" : "spot",
# Region speficiation
"region" : "us-west-2",
"availability_zone" : "us-west-2b",
# Machine type - instance type configuration.
"master_type" : "m4.2xlarge",
"worker_type" : "m4.2xlarge",
# please only use this AMI for pytorch
"image_id": "ami-f60aba8e",
# Launch specifications
"spot_price" : "0.15", # Has to be a string
# SSH configuration
"ssh_username" : "ubuntu", # For sshing. E.G: ssh ssh_username@hostname
"path_to_keyfile" : "/home/hwang/My_Code/AWS/HongyiScript.pem",
# NFS configuration
# To set up these values, go to Services > ElasticFileSystem > Create new filesystem, and follow the directions.
#"nfs_ip_address" : "172.31.3.173", # us-west-2c
#"nfs_ip_address" : "172.31.35.0", # us-west-2a
"nfs_ip_address" : "172.31.14.225", # us-west-2b
"nfs_mount_point" : "/home/ubuntu/shared", # NFS base dir
"base_out_dir" : "%(nfs_mount_point)s/%(name)s", # Master writes checkpoints to this directory. Outfiles are written to this directory.
"setup_commands" :
[
# "sudo rm -rf %(base_out_dir)s",
"mkdir %(base_out_dir)s",
],
# Command specification
# Master pre commands are run only by the master
"master_pre_commands" :
[
"cd my_mxnet",
"git fetch && git reset --hard origin/master",
"cd cifar10",
"ls",
# "cd distributed_tensorflow/DistributedResNet",
# "git fetch && git reset --hard origin/master",
],
# Pre commands are run on every machine before the actual training.
"pre_commands" :
[
"cd my_mxnet",
"git fetch && git reset --hard origin/master",
"cd cifar10",
],
# Model configuration
"batch_size" : "32",
"max_steps" : "2000",
"initial_learning_rate" : ".001",
"learning_rate_decay_factor" : ".95",
"num_epochs_per_decay" : "1.0",
# Train command specifies how the ps/workers execute tensorflow.
# PS_HOSTS - special string replaced with actual list of ps hosts.
# TASK_ID - special string replaced with actual task index.
# JOB_NAME - special string replaced with actual job name.
# WORKER_HOSTS - special string replaced with actual list of worker hosts
# ROLE_ID - special string replaced with machine's identity (E.G: master, worker0, worker1, ps, etc)
# %(...)s - Inserts self referential string value.
"train_commands" :
[
"echo ========= Start ==========="
],
})
def mxnet_ec2_run(argv, configuration):
client = boto3.client("ec2", region_name=configuration["region"])
ec2 = boto3.resource("ec2", region_name=configuration["region"])
def sleep_a_bit():
time.sleep(5)
def summarize_instances(instances):
instance_type_to_instance_map = {}
for instance in sorted(instances, key=lambda x:x.id):
typ = instance.instance_type
if typ not in instance_type_to_instance_map:
instance_type_to_instance_map[typ] = []
instance_type_to_instance_map[typ].append(instance)
for type in instance_type_to_instance_map:
print("Type\t", type)
for instance in instance_type_to_instance_map[type]:
print("instance\t", instance, "\t", instance.public_ip_address)
print
for k,v in instance_type_to_instance_map.items():
print("%s - %d running" % (k, len(v)))
return instance_type_to_instance_map
def summarize_idle_instances(argv):
print("Idle instances: (Idle = not running tensorflow)")
summarize_instances(get_idle_instances())
def summarize_running_instances(argv):
print("Running instances: ")
summarize_instances(ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}]))
# Terminate all request.
def terminate_all_requests():
spot_requests = client.describe_spot_instance_requests()
spot_request_ids = []
for spot_request in spot_requests["SpotInstanceRequests"]:
if spot_request["State"] != "cancelled" and spot_request["LaunchSpecification"]["KeyName"] == configuration["key_name"]:
spot_request_id = spot_request["SpotInstanceRequestId"]
spot_request_ids.append(spot_request_id)
if len(spot_request_ids) != 0:
print("Terminating spot requests: %s" % " ".join([str(x) for x in spot_request_ids]))
client.cancel_spot_instance_requests(SpotInstanceRequestIds=spot_request_ids)
# Wait until all are cancelled.
# TODO: Use waiter class
done = False
while not done:
print("Waiting for all spot requests to be terminated...")
done = True
spot_requests = client.describe_spot_instance_requests()
states = [x["State"] for x in spot_requests["SpotInstanceRequests"] if x["LaunchSpecification"]["KeyName"] == configuration["key_name"]]
for state in states:
if state != "cancelled":
done = False
sleep_a_bit()
# Terminate all instances in the configuration
# Note: all_instances = ec2.instances.all() to get all intances
def terminate_all_instances():
live_instances = ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}])
all_instance_ids = [x.id for x in live_instances]
print([x.id for x in live_instances])
if len(all_instance_ids) != 0:
print("Terminating instances: %s" % (" ".join([str(x) for x in all_instance_ids])))
client.terminate_instances(InstanceIds=all_instance_ids)
# Wait until all are terminated
# TODO: Use waiter class
done = False
while not done:
print("Waiting for all instances to be terminated...")
done = True
instances = ec2.instances.all()
for instance in instances:
if instance.state == "active":
done = False
sleep_a_bit()
# Launch instances as specified in the configuration.
def launch_instances():
method = "spot"
if "method" in configuration.keys():
method = configuration["method"]
worker_instance_type, worker_count = configuration["worker_type"], configuration["n_workers"]
master_instance_type, master_count = configuration["master_type"], configuration["n_masters"]
specs = [(worker_instance_type, worker_count),
(master_instance_type, master_count)]
for (instance_type, count) in specs:
launch_specs = {"KeyName" : configuration["key_name"],
"ImageId" : configuration["image_id"],
"InstanceType" : instance_type,
"Placement" : {"AvailabilityZone":configuration["availability_zone"]},
"SecurityGroups": ["default"]}
if method == "spot":
# TODO: EBS optimized? (Will incur extra hourly cost)
client.request_spot_instances(InstanceCount=count,
LaunchSpecification=launch_specs,
SpotPrice=configuration["spot_price"])
elif method == "reserved":
client.run_instances(ImageId=launch_specs["ImageId"],
MinCount=count,
MaxCount=count,
KeyName=launch_specs["KeyName"],
InstanceType=launch_specs["InstanceType"],
Placement=launch_specs["Placement"],
SecurityGroups=launch_specs["SecurityGroups"])
else:
print("Unknown method: %s" % method)
sys.exit(-1)
# TODO: use waiter class?
def wait_until_running_instances_initialized():
done = False
while not done:
print("Waiting for instances to be initialized...")
done = True
live_instances = ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}])
ids = [x.id for x in live_instances]
resps_list = [client.describe_instance_status(InstanceIds=ids[i:i+50]) for i in range(0, len(ids), 50)]
statuses = []
for resp in resps_list:
statuses += [x["InstanceStatus"]["Status"] for x in resp["InstanceStatuses"]]
#resps = client.describe_instance_status(InstanceIds=ids)
#for resp in resps["InstanceStatuses"]:
# if resp["InstanceStatus"]["Status"] != "ok":
# done = False
print(statuses)
done = statuses.count("ok") == len(statuses)
if len(ids) <= 0:
done = False
sleep_a_bit()
# Waits until status requests are all fulfilled.
# Prints out status of request in between time waits.
# TODO: Use waiter class
def wait_until_instance_request_status_fulfilled():
requests_fulfilled = False
n_active_or_open = 0
while not requests_fulfilled or n_active_or_open == 0:
requests_fulfilled = True
statuses = client.describe_spot_instance_requests()
print("InstanceRequestId, InstanceType, SpotPrice, State - Status : StatusMessage")
print("-------------------------------------------")
n_active_or_open = 0
for instance_request in statuses["SpotInstanceRequests"]:
if instance_request["LaunchSpecification"]["KeyName"] != configuration["key_name"]:
continue
sid = instance_request["SpotInstanceRequestId"]
machine_type = instance_request["LaunchSpecification"]["InstanceType"]
price = instance_request["SpotPrice"]
state = instance_request["State"]
status, status_string = instance_request["Status"]["Code"], instance_request["Status"]["Message"]
if state == "active" or state == "open":
n_active_or_open += 1
print("%s, %s, %s, %s - %s : %s" % (sid, machine_type, price, state, status, status_string))
if state != "active":
requests_fulfilled = False
print("-------------------------------------------")
sleep_a_bit()
# Create a client to the instance
def connect_client(instance):
client = pm.SSHClient()
host = instance.public_ip_address
client.set_missing_host_key_policy(pm.AutoAddPolicy())
client.connect(host, username=configuration["ssh_username"], key_filename=configuration["path_to_keyfile"])
return client
# Takes a list of commands (E.G: ["ls", "cd models"]
# and executes command on instance, returning the stdout.
# Executes everything in one session, and returns all output from all the commands.
def run_ssh_commands(instance, commands):
done = False
while not done:
try:
print("Instance %s, Running ssh commands:\n%s\n" % (instance.public_ip_address, "\n".join(commands)))
# Always need to exit
commands.append("exit")
# Set up ssh client
client = connect_client(instance)
# Clear the stdout from ssh'ing in
# For each command perform command and read stdout
commandstring = "\n".join(commands)
stdin, stdout, stderr = client.exec_command(commandstring)
output = stdout.read()
# Close down
stdout.close()
stdin.close()
client.close()
done = True
except Exception as e:
done = False
print(e.message)
return output
def run_ssh_commands_parallel(instance, commands, q):
output = run_ssh_commands(instance, commands)
q.put((instance, output))
# Checks whether instance is idle. Assumed that instance is up and running.
# An instance is idle if it is not running tensorflow...
# Returns a tuple of (instance, is_instance_idle). We return a tuple for multithreading ease.
def is_instance_idle(q, instance):
python_processes = run_ssh_commands(instance, ["ps aux | grep python"])
q.put((instance, not "ps_hosts" in python_processes and not "ps_workers" in python_processes))
# Idle instances are running instances that are not running the inception model.
# We check whether an instance is running the inception model by ssh'ing into a running machine,
# and checking whether python is running.
def get_idle_instances():
live_instances = ec2.instances.filter(
Filters=[{'Name': 'instance-state-name', 'Values': ['running']},
{'Name': 'key-name', 'Values': [configuration["key_name"]]}])
threads = []
q = Queue.Queue()
# Run commands in parallel, writing to the queue
for instance in live_instances:
t = threading.Thread(target=is_instance_idle, args=(q, instance))
t.daemon = True
t.start()
threads.append(t)
# Wait for threads to finish
for thread in threads:
thread.join()
# Collect idle instances
idle_instances = []
while not q.empty():
instance, is_idle = q.get()
if is_idle:
idle_instances.append(instance)
return idle_instances
def get_instance_requirements():
# Get the requirements given the specification of worker/master/etc machine types
worker_instance_type, worker_count = configuration["worker_type"], configuration["n_workers"]
master_instance_type, master_count = configuration["master_type"], configuration["n_masters"]
specs = [(worker_instance_type, worker_count),
(master_instance_type, master_count)]
reqs = {}
for (type_needed, count_needed) in specs:
if type_needed not in reqs:
reqs[type_needed] = 0
reqs[type_needed] += count_needed
return reqs
# Returns whether the idle instances satisfy the specs of the configuration.
def check_idle_instances_satisfy_configuration():
# Create a map of instance types to instances of that type
idle_instances = get_idle_instances()
instance_type_to_instance_map = summarize_instances(idle_instances)
# Get instance requirements
reqs = get_instance_requirements()
# Check the requirements are satisfied.
print("Checking whether # of running instances satisfies the configuration...")
for k,v in instance_type_to_instance_map.items():
n_required = 0 if k not in reqs else reqs[k]
print("%s - %d running vs %d required" % (k,len(v),n_required))
if len(v) < n_required:
print("Error, running instances failed to satisfy configuration requirements")
sys.exit(0)
print("Success, running instances satisfy configuration requirement")
def shut_everything_down(argv):
terminate_all_requests()
terminate_all_instances()
def run_mxnet_grid_search(argv, port=1334):
# Check idle instances satisfy configs
check_idle_instances_satisfy_configuration()
# Get idle instances
idle_instances = get_idle_instances()
# Assign instances for worker/ps/etc
instance_type_to_instance_map = summarize_instances(idle_instances)
specs = {
"master" : {"instance_type" : configuration["master_type"],
"n_required" : configuration["n_masters"]},
"worker" : {"instance_type" : configuration["worker_type"],
"n_required" : configuration["n_workers"]}
}
machine_assignments = {
"master" : [],
"worker" : []
}
for role, requirement in sorted(specs.items(), key=lambda x:x[0]):
instance_type_for_role = requirement["instance_type"]
n_instances_needed = requirement["n_required"]
instances_to_assign, rest = instance_type_to_instance_map[instance_type_for_role][:n_instances_needed], instance_type_to_instance_map[instance_type_for_role][n_instances_needed:]
instance_type_to_instance_map[instance_type_for_role] = rest
machine_assignments[role] = instances_to_assign
# Construct the host strings necessary for running the inception command.
# Note we use private ip addresses to avoid EC2 transfer costs.
worker_host_string = ",".join([x.private_ip_address+":"+str(port) for x in machine_assignments["master"] + machine_assignments["worker"]])
# Create a map of command&machine assignments
command_machine_assignments = {}
setup_machine_assignments = {}
# Construct the master command
command_machine_assignments["master"] = {"instance" : machine_assignments["master"][0], "commands" : list(configuration["master_pre_commands"])}
# setup_machine_assignments["master"] = {"instance" : machine_assignments["master"][0], "commands" : list(configuration["setup_commands"])}
for command_string in configuration["train_commands"]:
command_machine_assignments["master"]["commands"].append(command_string.replace("JOB_NAME", "worker").replace("WORKER_HOSTS", worker_host_string).replace("ROLE_ID", "master"))
print(command_machine_assignments)
# Construct the worker commands
for worker_id, instance in enumerate(machine_assignments["worker"]):
name = "worker_%d" % worker_id
command_machine_assignments[name] = {"instance" : instance,
"commands" : list(configuration["pre_commands"])}
for command_string in configuration["train_commands"]:
command_machine_assignments[name]["commands"].append(command_string.replace("TASK_ID", "%d" % (worker_id+1)).replace("JOB_NAME", "worker").replace("WORKER_HOSTS", worker_host_string).replace("ROLE_ID", name))
print(command_machine_assignments)
# Run the commands via ssh in parallel
threads = []
q = Queue.Queue()
for name, command_and_machine in setup_machine_assignments.items():
instance = command_and_machine["instance"]
commands = command_and_machine["commands"]
print("-----------------------")
print("Pre Command: %s\n" % " ".join(commands))
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
# Wait until commands are all finished
for t in threads:
t.join()
threads = []
q = Queue.Queue()
running_process = 0
for name, command_and_machine in command_machine_assignments.items():
instance = command_and_machine["instance"]
neo_commands = "python train_cifar10.py --running_mode=grid_search --gpus=0 "\
"--running_process={} "\
"--batch-size={} "\
"--dir={}/grid_search> {}/grid_search/batch_size_{}/running_{}_process.out 2>&1 &".format(
running_process,
configuration['batch_size'],
configuration['nfs_mount_point'],
configuration['nfs_mount_point'],
configuration['batch_size'],
running_process)
commands = command_and_machine["commands"]
commands.append('mkdir {}/grid_search'.format(configuration['nfs_mount_point']))
commands.append('mkdir {}/grid_search/batch_size_{}'.format(
configuration['nfs_mount_point'],
configuration['batch_size']))
commands.append(neo_commands)
print("-----------------------")
print("Command: %s\n" % " ".join(commands))
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
running_process += 1
# Wait until commands are all finished
for t in threads:
t.join()
# Print the output
while not q.empty():
instance, output = q.get()
print(instance.public_ip_address)
print(output)
# Debug print
instances = []
print("\n--------------------------------------------------\n")
print("Machine assignments:")
print("------------------------")
for name, command_and_machine in command_machine_assignments.items():
instance = command_and_machine["instance"]
instances.append(instance)
commands = command_and_machine["commands"]
ssh_command = "ssh -i %s %s@%s" % (configuration["path_to_keyfile"], configuration["ssh_username"], instance.public_ip_address)
print("%s - %s" % (name, instance.instance_id))
print("To ssh: %s" % ssh_command)
print("------------------------")
# Print out list of instance ids (which will be useful in selctively stopping inception
# for given instances.
instance_cluster_string = ",".join([x.instance_id for x in instances])
print("\nInstances cluster string: %s" % instance_cluster_string)
# Print out the id of the configuration file
cluster_save = {
"configuration" : configuration,
"name" : configuration["name"],
"command_machine_assignments" : command_machine_assignments,
"cluster_string" : instance_cluster_string
}
return cluster_save
def run_mxnet_loss_curve(argv, port=1334):
# Check idle instances satisfy configs
check_idle_instances_satisfy_configuration()
# Get idle instances
idle_instances = get_idle_instances()
# Assign instances for worker/ps/etc
instance_type_to_instance_map = summarize_instances(idle_instances)
specs = {
"master" : {"instance_type" : configuration["master_type"],
"n_required" : configuration["n_masters"]},
"worker" : {"instance_type" : configuration["worker_type"],
"n_required" : configuration["n_workers"]}
}
machine_assignments = {
"master" : [],
"worker" : []
}
for role, requirement in sorted(specs.items(), key=lambda x:x[0]):
instance_type_for_role = requirement["instance_type"]
n_instances_needed = requirement["n_required"]
instances_to_assign, rest = instance_type_to_instance_map[instance_type_for_role][:n_instances_needed], instance_type_to_instance_map[instance_type_for_role][n_instances_needed:]
instance_type_to_instance_map[instance_type_for_role] = rest
machine_assignments[role] = instances_to_assign
# Construct the host strings necessary for running the inception command.
# Note we use private ip addresses to avoid EC2 transfer costs.
worker_host_string = ",".join([x.private_ip_address+":"+str(port) for x in machine_assignments["master"] + machine_assignments["worker"]])
# Create a map of command&machine assignments
command_machine_assignments = {}
setup_machine_assignments = {}
# Construct the master command
command_machine_assignments["master"] = {"instance" : machine_assignments["master"][0], "commands" : list(configuration["master_pre_commands"])}
# setup_machine_assignments["master"] = {"instance" : machine_assignments["master"][0], "commands" : list(configuration["setup_commands"])}
for command_string in configuration["train_commands"]:
command_machine_assignments["master"]["commands"].append(command_string.replace("JOB_NAME", "worker").replace("WORKER_HOSTS", worker_host_string).replace("ROLE_ID", "master"))
print(command_machine_assignments)
# Construct the worker commands
for worker_id, instance in enumerate(machine_assignments["worker"]):
name = "worker_%d" % worker_id
command_machine_assignments[name] = {"instance" : instance,
"commands" : list(configuration["pre_commands"])}
for command_string in configuration["train_commands"]:
command_machine_assignments[name]["commands"].append(command_string.replace("TASK_ID", "%d" % (worker_id+1)).replace("JOB_NAME", "worker").replace("WORKER_HOSTS", worker_host_string).replace("ROLE_ID", name))
print(command_machine_assignments)
# Run the commands via ssh in parallel
threads = []
q = Queue.Queue()
for name, command_and_machine in setup_machine_assignments.items():
instance = command_and_machine["instance"]
commands = command_and_machine["commands"]
print("-----------------------")
print("Pre Command: %s\n" % " ".join(commands))
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
# Wait until commands are all finished
for t in threads:
t.join()
threads = []
q = Queue.Queue()
batch_size_list = [4, 32, 50, 100, 500, 1000]
learning_rate_list = [0.046, 0.05, 0.068, 0.068, 0.048, 0.086]
running_process = 0
for name, command_and_machine in command_machine_assignments.items():
instance = command_and_machine["instance"]
neo_commands = "python train_cifar10.py --running_mode=training --gpus=0 "\
"--batch-size={} "\
"--lr={} "\
"--model-prefix={}/model_checkpoints/batch_size_{} "\
"--dir={}/loss_curve > "\
"{}/loss_curve/running_batch_size_{}.out 2>&1 &".format(
batch_size_list[running_process],
learning_rate_list[running_process],
configuration['nfs_mount_point'],
batch_size_list[running_process],
configuration['nfs_mount_point'],
configuration['nfs_mount_point'],
batch_size_list[running_process])
commands = command_and_machine["commands"]
commands.append('mkdir {}/model_checkpoints/'.format(configuration['nfs_mount_point']))
commands.append('mkdir {}/loss_curve'.format(configuration['nfs_mount_point']))
commands.append(neo_commands)
print("-----------------------")
print("Command: %s\n" % " ".join(commands))
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
running_process += 1
# Wait until commands are all finished
for t in threads:
t.join()
# Print the output
while not q.empty():
instance, output = q.get()
print(instance.public_ip_address)
print(output)
# Debug print
instances = []
print("\n--------------------------------------------------\n")
print("Machine assignments:")
print("------------------------")
for name, command_and_machine in command_machine_assignments.items():
instance = command_and_machine["instance"]
instances.append(instance)
commands = command_and_machine["commands"]
ssh_command = "ssh -i %s %s@%s" % (configuration["path_to_keyfile"], configuration["ssh_username"], instance.public_ip_address)
print("%s - %s" % (name, instance.instance_id))
print("To ssh: %s" % ssh_command)
print("------------------------")
# Print out list of instance ids (which will be useful in selctively stopping inception
# for given instances.
instance_cluster_string = ",".join([x.instance_id for x in instances])
print("\nInstances cluster string: %s" % instance_cluster_string)
# Print out the id of the configuration file
cluster_save = {
"configuration" : configuration,
"name" : configuration["name"],
"command_machine_assignments" : command_machine_assignments,
"cluster_string" : instance_cluster_string
}
return cluster_save
def get_hosts(argv, port=22):
# Check idle instances satisfy configs
check_idle_instances_satisfy_configuration()
# Get idle instances
idle_instances = get_idle_instances()
# Assign instances for worker/ps/etc
instance_type_to_instance_map = summarize_instances(idle_instances)
specs = {
"master" : {"instance_type" : configuration["master_type"],
"n_required" : configuration["n_masters"]},
"worker" : {"instance_type" : configuration["worker_type"],
"n_required" : configuration["n_workers"]}
}
machine_assignments = {
"master" : [],
"worker" : []
}
for role, requirement in sorted(specs.items(), key=lambda x:x[0]):
instance_type_for_role = requirement["instance_type"]
n_instances_needed = requirement["n_required"]
instances_to_assign, rest = instance_type_to_instance_map[instance_type_for_role][:n_instances_needed], instance_type_to_instance_map[instance_type_for_role][n_instances_needed:]
instance_type_to_instance_map[instance_type_for_role] = rest
machine_assignments[role] = instances_to_assign
# Construct the host strings necessary for running the inception command.
# Note we use private ip addresses to avoid EC2 transfer costs.
worker_host_string = ",".join([x.private_ip_address+":"+str(port) for x in machine_assignments["master"] + machine_assignments["worker"]])
hosts_out = open('hosts', 'w')
print('master ip ', machine_assignments['master'][0].public_ip_address)
count = 0
for instance in machine_assignments["master"] + machine_assignments["worker"]:
count += 1
print('{}\tdeeplearning-worker{}'.format(instance.private_ip_address, count), end='\n', file=hosts_out)
hosts_out.flush()
hosts_out.close()
hosts_alias_out = open('hosts_alias', 'w')
count = 0
for _ in machine_assignments["master"] + machine_assignments["worker"]:
count += 1
print('deeplearning-worker{}'.format(count), end='\n', file=hosts_alias_out)
hosts_alias_out.flush()
hosts_alias_out.close()
hosts_alias_out = open('hosts_address', 'w')
count = 0
for instance in machine_assignments["master"] + machine_assignments["worker"]:
count += 1
print('{}'.format(instance.private_ip_address), end='\n', file=hosts_alias_out)
hosts_alias_out.flush()
hosts_alias_out.close()
# # Create a map of command&machine assignments
# command_machine_assignments = {}
# setup_machine_assignments = {}
#
# # Construct the master command
# command_machine_assignments["master"] = {"instance" : machine_assignments["master"][0], "commands" : list(configuration["master_pre_commands"])}
# for command_string in configuration["train_commands"]:
# command_machine_assignments["master"]["commands"].append(command_string.replace("JOB_NAME", "worker").replace("WORKER_HOSTS", worker_host_string).replace("ROLE_ID", "master"))
# print(command_machine_assignments)
#
# # Construct the worker commands
# for worker_id, instance in enumerate(machine_assignments["worker"]):
# name = "worker_%d" % worker_id
# command_machine_assignments[name] = {"instance" : instance,
# "commands" : list(configuration["pre_commands"])}
# for command_string in configuration["train_commands"]:
# command_machine_assignments[name]["commands"].append(command_string.replace("TASK_ID", "%d" % (worker_id+1)).replace("JOB_NAME", "worker").replace("WORKER_HOSTS", worker_host_string).replace("ROLE_ID", name))
#
# print(command_machine_assignments)
#
# # Run the commands via ssh in parallel
# threads = []
# q = Queue.Queue()
#
# for name, command_and_machine in setup_machine_assignments.items():
# instance = command_and_machine["instance"]
# commands = command_and_machine["commands"]
# print("-----------------------")
# print("Pre Command: %s\n" % " ".join(commands))
# t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
# t.start()
# threads.append(t)
#
# # Wait until commands are all finished
# for t in threads:
# t.join()
#
# threads = []
# q = Queue.Queue()
#
# batch_size_list = [4, 32, 50, 100, 500, 1000]
# learning_rate_list = [0.046, 0.05, 0.068, 0.068, 0.048, 0.086]
# running_process = 0
# for name, command_and_machine in command_machine_assignments.items():
# instance = command_and_machine["instance"]
# neo_commands = "python train_cifar10.py --running_mode=training --gpus=0 "\
# "--batch-size={} "\
# "--lr={} "\
# "--model-prefix={}/model_checkpoints/batch_size_{} "\
# "--dir={}/loss_curve > "\
# "{}/loss_curve/running_batch_size_{}.out 2>&1 &".format(
# batch_size_list[running_process],
# learning_rate_list[running_process],
# configuration['nfs_mount_point'],
# batch_size_list[running_process],
# configuration['nfs_mount_point'],
# configuration['nfs_mount_point'],
# batch_size_list[running_process])
#
# commands = command_and_machine["commands"]
# commands.append('mkdir {}/model_checkpoints/'.format(configuration['nfs_mount_point']))
# commands.append('mkdir {}/loss_curve'.format(configuration['nfs_mount_point']))
# commands.append(neo_commands)
#
# print("-----------------------")
# print("Command: %s\n" % " ".join(commands))
# t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
# t.start()
# threads.append(t)
# running_process += 1
#
# # Wait until commands are all finished
# for t in threads:
# t.join()
#
# # Print the output
# while not q.empty():
# instance, output = q.get()
# print(instance.public_ip_address)
# print(output)
#
# # Debug print
# instances = []
# print("\n--------------------------------------------------\n")
# print("Machine assignments:")
# print("------------------------")
# for name, command_and_machine in command_machine_assignments.items():
# instance = command_and_machine["instance"]
# instances.append(instance)
# commands = command_and_machine["commands"]
# ssh_command = "ssh -i %s %s@%s" % (configuration["path_to_keyfile"], configuration["ssh_username"], instance.public_ip_address)
# print("%s - %s" % (name, instance.instance_id))
# print("To ssh: %s" % ssh_command)
# print("------------------------")
#
# # Print out list of instance ids (which will be useful in selctively stopping inception
# # for given instances.
# instance_cluster_string = ",".join([x.instance_id for x in instances])
# print("\nInstances cluster string: %s" % instance_cluster_string)
#
# # Print out the id of the configuration file
# cluster_save = {
# "configuration" : configuration,
# "name" : configuration["name"],
# "command_machine_assignments" : command_machine_assignments,
# "cluster_string" : instance_cluster_string
# }
#
# return cluster_save
return
def kill_python(argv):
if len(argv) != 3:
print("Usage: python inception_ec2.py kill_python instance_id1,instance_id2,id3...")
sys.exit(0)
cluster_instance_string = argv[2]
instance_ids_to_shutdown = cluster_instance_string.split(",")
live_instances = ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}])
threads = []
q = Queue.Queue()
for instance in live_instances:
if instance.instance_id in instance_ids_to_shutdown:
commands = ["sudo pkill -9 python"]
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
for thread in threads:
thread.join()
summarize_idle_instances(None)
def kill_all_python(argv):
live_instances = ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}])
threads = []
q = Queue.Queue()
for instance in live_instances:
commands = ["sudo pkill -9 python"]
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
for thread in threads:
thread.join()
summarize_idle_instances(None)
def run_command(argv, quiet=False):
if len(argv) != 4:
print("Usage: python inception_ec2.py run_command instance_id1,instance_id2,id3... command")
sys.exit(0)
cluster_instance_string = argv[2]
command = argv[3]
instance_ids_to_run_command = cluster_instance_string.split(",")
live_instances = ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}])
threads = []
q = Queue.Queue()
for instance in live_instances:
if instance.instance_id in instance_ids_to_run_command:
commands = [command]
t = threading.Thread(target=run_ssh_commands_parallel, args=(instance, commands, q))
t.start()
threads.append(t)
for thread in threads:
thread.join()
while not q.empty():
instance, output = q.get()
if not quiet:
print(instance, output)
# Setup nfs on all instances
def setup_nfs():
print("Clearing previous nfs file system...")
live_instances = ec2.instances.filter(Filters=[{'Name': 'instance-state-name', 'Values': ['running']}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}, {'Name': 'key-name', 'Values': [configuration["key_name"]]}])
live_instances_string = ",".join([x.instance_id for x in live_instances])
rm_command = "sudo rm -rf %s" % configuration["nfs_mount_point"]
argv = ["python", "inception_ec2.py", live_instances_string, rm_command]
# argv = ["python", "inception_ec2.py", live_instances_string]
run_command(argv, quiet=True)
print("Installing nfs on all running instances...")
update_command = "sudo apt-get -y update"
install_nfs_command = "sudo apt-get -y install nfs-common"
create_mount_command = "mkdir %s" % configuration["nfs_mount_point"]
setup_nfs_command = "sudo mount -t nfs4 -o nfsvers=4.1,rsize=1048576,wsize=1048576,hard,timeo=600,retrans=2 %s:/ %s" % (configuration["nfs_ip_address"], configuration["nfs_mount_point"])
reduce_permissions_command = "sudo chmod 777 %s " % configuration["nfs_mount_point"]
command = update_command + " && " + install_nfs_command + " && " + create_mount_command + " && " + setup_nfs_command + " && " + reduce_permissions_command
# pretty hackish
argv = ["python", "inception_ec2.py", live_instances_string, command]
run_command(argv, quiet=True)
return
# Launch instances as specified by the configuration.
# We also want a shared filesystem to write model checkpoints.
# For simplicity we will have the user specify the filesystem via the config.
def launch(argv):
method = "spot"
if "method" in configuration:
method = configuration["method"]
launch_instances()
if method == "spot":
wait_until_instance_request_status_fulfilled()
wait_until_running_instances_initialized()
print('setup nfs')
setup_nfs()
def clean_launch_and_run(argv):
# 1. Kills all instances in region
# 2. Kills all requests in region
# 3. Launches requests
# 5. Waits until launch requests have all been satisfied,
# printing status outputs in the meanwhile
# 4. Checks that configuration has been satisfied
# 5. Runs inception
shut_everything_down(None)
launch(None)
return run_mxnet_grid_search(None)
def help(hmap):
print("Usage: python inception_ec2.py [command]")
print("Commands:")
for k,v in hmap.items():
print("%s - %s" % (k,v))
##############################
# tf_ec2 main starting point #
##############################
command_map = {
"launch" : launch,
"clean_launch_and_run" : clean_launch_and_run,
"shutdown" : shut_everything_down,
"run_mxnet_grid_search": run_mxnet_grid_search,
"run_mxnet_loss_curve": run_mxnet_loss_curve,
"get_hosts": get_hosts,
"kill_all_python" : kill_all_python,
"list_idle_instances" : summarize_idle_instances,
"list_running_instances" : summarize_running_instances,
"kill_python" : kill_python,
"run_command" : run_command,
"setup_nfs": setup_nfs,
}
help_map = {
"launch" : "Launch instances",
"clean_launch_and_run" : "Shut everything down, launch instances, wait until requests fulfilled, check that configuration is fulfilled, and launch and run inception.",
"shutdown" : "Shut everything down by cancelling all instance requests, and terminating all instances.",
"list_idle_instances" : "Lists all idle instances. Idle instances are running instances not running tensorflow.",
"list_running_instances" : "Lists all running instances.",
"run_mxnet_grid_search": "",
"run_mxnet_loss_curve": "",
"setup_nfs": "",
"kill_all_python" : "Kills python running inception training on ALL instances.",
"kill_python" : "Kills python running inception on instances indicated by instance id string separated by ',' (no spaces).",
"run_command" : "Runs given command on instances selcted by instance id string, separated by ','.",
}
if len(argv) < 2:
help(help_map)
sys.exit(0)
command = argv[1]
return command_map[command](argv)
if __name__ == "__main__":
print(cfg)
mxnet_ec2_run(sys.argv, cfg)