-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
executable file
·494 lines (441 loc) · 25.9 KB
/
main.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
import math
import time
import json
import sys
from algorithm.cMMAC import *
from algorithm.GPG import *
from env.platform import *
from env.env_run import *
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
def flatten(list):
return [y for x in list for y in x]
def calculate_reward(master1, master2, cur_done, cur_undone):
weight = 1.0
all_task = [float(cur_done[0] + cur_undone[0]), float(cur_done[1] + cur_undone[1])]
fail_task = [float(cur_undone[0]), float(cur_undone[1])]
reward = []
# The ratio of requests that violate delay requirements
task_fail_rate = []
if all_task[0] != 0:
task_fail_rate.append(fail_task[0] / all_task[0])
else:
task_fail_rate.append(0)
if all_task[1] != 0:
task_fail_rate.append(fail_task[1] / all_task[1])
else:
task_fail_rate.append(0)
# The standard deviation of the CPU and memory usage
standard_list = []
use_rate1 = []
use_rate2 = []
for i in range(3):
use_rate1.append(master1.node_list[i].cpu / master1.node_list[i].cpu_max)
use_rate1.append(master1.node_list[i].mem / master1.node_list[i].mem_max)
use_rate2.append(master2.node_list[i].cpu / master2.node_list[i].cpu_max)
use_rate2.append(master2.node_list[i].mem / master2.node_list[i].mem_max)
standard_list.append(np.std(use_rate1, ddof=1))
standard_list.append(np.std(use_rate2, ddof=1))
reward.append(math.exp(-task_fail_rate[0]) + weight * math.exp(-standard_list[0]))
reward.append(math.exp(-task_fail_rate[1]) + weight * math.exp(-standard_list[1]))
return reward
def to_grid_rewards(node_reward):
return np.array(node_reward).reshape([-1, 1])
def execution(RUN_TIMES, BREAK_POINT, TRAIN_TIMES, CHO_CYCLE):
############ Set up according to your own needs ###########
# The parameters are set to support the operation of the program, and may not be consistent with the actual system
vaild_node = 6 # Number of edge nodes available
SLOT_TIME = 0.5 # Time of one slot
MAX_TESK_TYPE = 12 # Number of tesk types， 任务类总数
POD_CPU = 15.0 # CPU resources required for a POD
POD_MEM = 1.0 # Memory resources required for a POD
# Resource demand coefficients for different types of services
service_coefficient = [0.8, 0.8, 0.9, 0.9, 1.0, 1.0, 1.1, 1.1, 1.2, 1.2, 1.3, 1.3, 1.4, 1.4]
# Parameters related to DRL
epsilon = 0.5
gamma = 0.9
learning_rate = 1e-3
action_dim = 7 #6个节点+一个云端
state_dim = 88 #每个eap agent的总共state
node_input_dim = 24 #可以认为是没个edge node的维数
cluster_input_dim = 24 #summary
hid_dims = [16, 8]
output_dim = 8 #这里务必留个心眼，默认为图网络输出的embedding
max_depth = 8
entropy_weight_init = 1
exec_cap = 24
entropy_weight_min = 0.0001
entropy_weight_decay = 1e-3
# Parameters related to GPU
worker_num_gpu = 0
worker_gpu_fraction = 0.1
#####################################################################
########### Init ###########
record = []
throughput_list = []
sum_rewards = []
achieve_num = []
achieve_num_sum = []
fail_num = []
deploy_reward = []
current_time = str(time.time())
log_dir = "./log/{}/".format(current_time)
all_rewards = []
order_response_rate_episode = []
episode_rewards = []
record_all_order_response_rate = []
sess = tf.Session()
tf.set_random_seed(1)
q_estimator = Estimator(sess, action_dim, state_dim, 2, scope="q_estimator", summaries_dir=log_dir)
sess.run(tf.global_variables_initializer())
replay = ReplayMemory(memory_size=1e+6, batch_size=int(3e+3))
policy_replay = policyReplayMemory(memory_size=1e+6, batch_size=int(3e+3))
saver = tf.compat.v1.train.Saver()
global_step1 = 0
global_step2 = 0
all_task1 = get_all_task('./data/Task_1.csv')
all_task2 = get_all_task('./data/Task_2.csv')
config = tf.ConfigProto(device_count={'GPU': worker_num_gpu},
gpu_options=tf.GPUOptions(per_process_gpu_memory_fraction=worker_gpu_fraction))
sess = tf.Session(config=config)
orchestrate_agent = OrchestrateAgent(sess, node_input_dim, cluster_input_dim, hid_dims, output_dim, max_depth,
range(1, exec_cap + 1))
exp = {'node_inputs': [], 'cluster_inputs': [], 'reward': [], 'wall_time': [], 'node_act_vec': [],
'cluster_act_vec': []}
for n_iter in np.arange(RUN_TIMES): #相当于linspace的作用，大时间
########### Initialize the setup and repeat the experiment many times ###########
batch_reward = []
cur_time = 0
entropy_weight = entropy_weight_init
order_response_rates = []
pre_done = [0, 0]
pre_undone = [0, 0]
context = [1, 1]
############ Set up according to your own needs ###########
# The parameters here are set only to support the operation of the program, and may not be consistent with the actual system
deploy_state = [[0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 1],
[0, 0, 0, 1, 0, 0, 1, 1, 0, 1, 0, 0], [0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 0, 1],
[0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1]]
# Create clusters based on the hardware resources you need
# 一共6个node
node1_1 = Node(100.0, 4.0, [], []) # (cpu, mem,...)
node1_2 = Node(200.0, 6.0, [], [])
node1_3 = Node(100.0, 8.0, [], [])
node_list1 = [node1_1, node1_2, node1_3]
node2_1 = Node(200.0, 8.0, [], [])
node2_2 = Node(100.0, 2.0, [], [])
node2_3 = Node(200.0, 6.0, [], [])
node_list2 = [node2_1, node2_2, node2_3]
# (cpu, mem,..., achieve task num, give up task num)
#Master就是eap,或者感觉像集群
master1 = Master(200.0, 8.0, node_list1, [], all_task1, 0, 0, 0, [0] * MAX_TESK_TYPE, [0] * MAX_TESK_TYPE)
master2 = Master(200.0, 8.0, node_list2, [], all_task2, 0, 0, 0, [0] * MAX_TESK_TYPE, [0] * MAX_TESK_TYPE)
cloud = Cloud([], [], sys.maxsize, sys.maxsize) # (..., cpu, mem)
################################################################################################
for i in range(MAX_TESK_TYPE):
docker = Docker(POD_MEM * service_coefficient[i], POD_CPU * service_coefficient[i], cur_time, i, [-1])
cloud.service_list.append(docker)
# Crerate dockers based on deploy_state
#######################################################
###这段以文有点大
for i in range(vaild_node):
for ii in range(MAX_TESK_TYPE):
dicision = deploy_state[i][ii]
if i < 3 and dicision == 1:
j = i
if master1.node_list[j].mem >= POD_MEM * service_coefficient[ii]:
docker = Docker(POD_MEM * service_coefficient[ii], POD_CPU * service_coefficient[ii], cur_time,
ii, [-1])
master1.node_list[j].mem = master1.node_list[j].mem - POD_MEM * service_coefficient[ii]
master1.node_list[j].service_list.append(docker)
if i >= 3 and dicision == 1:
j = i - 3
if master2.node_list[j].mem >= POD_MEM * service_coefficient[ii]:
docker = Docker(POD_MEM * service_coefficient[ii], POD_CPU * service_coefficient[ii], cur_time,
ii, [-1])
master2.node_list[j].mem = master2.node_list[j].mem - POD_MEM * service_coefficient[ii]
master2.node_list[j].service_list.append(docker)
###################################################################################
########### Each slot ########### 这里开始小时间
for slot in range(BREAK_POINT):
cur_time = cur_time + SLOT_TIME
########### Each frame ###########
if slot % CHO_CYCLE == 0 and slot != 0:
done_tasks = []
undone_tasks = []
curr_tasks_in_queue = []
# Get task state, include successful, failed, and unresolved
for i in range(MAX_TESK_TYPE):
done_tasks.append(float(master1.done_kind[i] + master2.done_kind[i]))
undone_tasks.append(float(master1.undone_kind[i] + master2.undone_kind[i]))
for i in range(3):
tmp = [0.0] * MAX_TESK_TYPE
for j in range(len(master1.node_list[i].task_queue)):
tmp[master1.node_list[i].task_queue[j][0]] = tmp[master1.node_list[i].task_queue[j][0]] + 1.0
curr_tasks_in_queue.append(tmp)
for i in range(3):
tmp = [0.0] * MAX_TESK_TYPE
for k in range(len(master2.node_list[i].task_queue)):
tmp[master2.node_list[i].task_queue[k][0]] = tmp[master2.node_list[i].task_queue[k][0]] + 1
curr_tasks_in_queue.append(tmp)
if slot != CHO_CYCLE:
exp['reward'].append(float(sum(deploy_reward)) / float(len(deploy_reward)))
deploy_reward = []
exp['wall_time'].append(cur_time)
deploy_state_float = []
for i in range(len(deploy_state)):
tmp = []
for j in range(len(deploy_state[0])):
tmp.append(float(deploy_state[i][j]))
deploy_state_float.append(tmp)
# Make decision of orchestration
change_node, change_service, exp = act_offload_agent(orchestrate_agent, exp, done_tasks,
undone_tasks, curr_tasks_in_queue,
deploy_state_float)
# print(change_node) #一共6个可用node
# print(change_service) #这个
# Execute orchestration
for i in range(len(change_node)):
if change_service[i] < 0:
# Delete docker and free memory
service_index = -1 * change_service[i] - 1
if change_node[i] < 3:
docker_idx = 0
while docker_idx < len(master1.node_list[change_node[i]].service_list):
if docker_idx >= len(master1.node_list[change_node[i]].service_list):
break
if master1.node_list[change_node[i]].service_list[docker_idx].kind == service_index:
master1.node_list[change_node[i]].mem = master1.node_list[change_node[i]].mem + \
master1.node_list[
change_node[i]].service_list[
docker_idx].mem
del master1.node_list[change_node[i]].service_list[docker_idx]
deploy_state[change_node[i]][service_index] = deploy_state[change_node[i]][
service_index] - 1.0
else:
docker_idx = docker_idx + 1
else:
node_index = change_node[i] - 3
docker_idx = 0
while docker_idx < len(master2.node_list[node_index].service_list):
if docker_idx >= len(master2.node_list[node_index].service_list):
break
if master2.node_list[node_index].service_list[docker_idx].kind == service_index:
master2.node_list[node_index].mem = master2.node_list[node_index].mem + \
master2.node_list[node_index].service_list[
docker_idx].mem
del master2.node_list[node_index].service_list[docker_idx]
deploy_state[node_index][service_index] = deploy_state[node_index][
service_index] - 1.0
else:
docker_idx = docker_idx + 1
else:
# Add docker and tack up memory
service_index = change_service[i] - 1
if change_node[i] < 3:
if master1.node_list[change_node[i]].mem >= POD_MEM * service_coefficient[service_index]:
docker = Docker(POD_MEM * service_coefficient[service_index],
POD_CPU * service_coefficient[service_index],
cur_time, service_index, [-1])
master1.node_list[change_node[i]].mem = master1.node_list[
change_node[i]].mem - POD_MEM * \
service_coefficient[service_index]
master1.node_list[change_node[i]].service_list.append(docker)
deploy_state[change_node[i]][service_index] = deploy_state[change_node[i]][
service_index] + 1
else:
node_index = change_node[i] - 3
if master2.node_list[node_index].mem >= POD_MEM * service_coefficient[service_index]:
docker = Docker(POD_MEM * service_coefficient[service_index],
POD_CPU * service_coefficient[service_index],
cur_time, service_index, [-1])
master2.node_list[node_index].mem = master2.node_list[node_index].mem - POD_MEM * \
service_coefficient[service_index]
master2.node_list[node_index].service_list.append(docker)
deploy_state[node_index][service_index] = deploy_state[node_index][service_index] + 1
# Save data
if slot > 3 * CHO_CYCLE:
exp_tmp = exp
del exp_tmp['node_inputs'][-1]
del exp_tmp['cluster_inputs'][-1]
del exp_tmp['node_act_vec'][-1]
del exp_tmp['cluster_act_vec'][-1]
entropy_weight, loss = train_orchestrate_agent(orchestrate_agent, exp_tmp, entropy_weight,
entropy_weight_min, entropy_weight_decay)
entropy_weight = decrease_var(entropy_weight,
entropy_weight_min, entropy_weight_decay)
# Get current task
master1 = update_task_queue(master1, cur_time, 0)
master2 = update_task_queue(master2, cur_time, 1)
task1 = [-1]
task2 = [-1]
if len(master1.task_queue) != 0:
task1 = master1.task_queue[0]
del master1.task_queue[0]
if len(master2.task_queue) != 0:
task2 = master2.task_queue[0]
del master2.task_queue[0]
curr_task = [task1, task2]
ava_node = []
for i in range(len(curr_task)):
tmp_list = [6] # Cloud computing
for ii in range(len(deploy_state)):
if deploy_state[ii][curr_task[i][0]] == 1:
tmp_list.append(ii)
ava_node.append(tmp_list)
# Current state of CPU and memory
cpu_list1 = []
mem_list1 = []
cpu_list2 = []
mem_list2 = []
task_num1 = [len(master1.task_queue)]
task_num2 = [len(master2.task_queue)]
for i in range(3):
cpu_list1.append([master1.node_list[i].cpu, master1.node_list[i].cpu_max])
mem_list1.append([master1.node_list[i].mem, master1.node_list[i].mem_max])
task_num1.append(len(master1.node_list[i].task_queue))
for i in range(3):
cpu_list2.append([master2.node_list[i].cpu, master2.node_list[i].cpu_max])
mem_list2.append([master2.node_list[i].mem, master2.node_list[i].mem_max])
task_num2.append(len(master2.node_list[i].task_queue))
s_grid = np.array([flatten(flatten([deploy_state, [task_num1], cpu_list1, mem_list1])),
flatten(flatten([deploy_state, [task_num2], cpu_list1, mem_list1]))])
# Dispatch decision
act, valid_action_prob_mat, policy_state, action_choosen_mat, \
curr_state_value, curr_neighbor_mask, next_state_ids = q_estimator.action(s_grid, ava_node, context,
epsilon)
# Put the current task on the queue based on dispatch decision
for i in range(len(act)):
if curr_task[i][0] == -1:
continue
if act[i] == 6:
cloud.task_queue.append(curr_task[i])
continue
if act[i] >= 0 and act[i] < 3:
master1.node_list[act[i]].task_queue.append(curr_task[i])
continue
if act[i] >= 3 and act[i] < 6:
master2.node_list[act[i] - 3].task_queue.append(curr_task[i])
continue
else:
pass
# Update state of task
for i in range(3):
master1.node_list[i].task_queue, undone, undone_kind = check_queue(master1.node_list[i].task_queue,
cur_time)
for j in undone_kind:
master1.undone_kind[j] = master1.undone_kind[j] + 1
master1.undone = master1.undone + undone[0]
master2.undone = master2.undone + undone[1]
master2.node_list[i].task_queue, undone, undone_kind = check_queue(master2.node_list[i].task_queue,
cur_time)
for j in undone_kind:
master2.undone_kind[j] = master2.undone_kind[j] + 1
master1.undone = master1.undone + undone[0]
master2.undone = master2.undone + undone[1]
cloud.task_queue, undone, undone_kind = check_queue(cloud.task_queue, cur_time)
master1.undone = master1.undone + undone[0]
master2.undone = master2.undone + undone[1]
# Update state of dockers in every node
for i in range(3):
master1.node_list[i], undone, done, done_kind, undone_kind = update_docker(master1.node_list[i],
cur_time,
service_coefficient, POD_CPU)
for j in range(len(done_kind)):
master1.done_kind[done_kind[j]] = master1.done_kind[done_kind[j]] + 1
for j in range(len(undone_kind)):
master1.undone_kind[undone_kind[j]] = master1.undone_kind[undone_kind[j]] + 1
master1.undone = master1.undone + undone[0]
master2.undone = master2.undone + undone[1]
master1.done = master1.done + done[0]
master2.done = master2.done + done[1]
master2.node_list[i], undone, done, done_kind, undone_kind = update_docker(master2.node_list[i],
cur_time,
service_coefficient, POD_CPU)
for j in range(len(done_kind)):
master1.done_kind[done_kind[j]] = master1.done_kind[done_kind[j]] + 1
for j in range(len(undone_kind)):
master1.undone_kind[undone_kind[j]] = master1.undone_kind[undone_kind[j]] + 1
master1.undone = master1.undone + undone[0]
master2.undone = master2.undone + undone[1]
master1.done = master1.done + done[0]
master2.done = master2.done + done[1]
cloud, undone, done, done_kind, undone_kind = update_docker(cloud, cur_time, service_coefficient, POD_CPU)
master1.undone = master1.undone + undone[0]
master2.undone = master2.undone + undone[1]
master1.done = master1.done + done[0]
master2.done = master2.done + done[1]
cur_done = [master1.done - pre_done[0], master2.done - pre_done[1]]
cur_undone = [master1.undone - pre_undone[0], master2.undone - pre_undone[1]]
pre_done = [master1.done, master2.done]
pre_undone = [master1.undone, master2.undone]
achieve_num.append(sum(cur_done))
fail_num.append(sum(cur_undone))
immediate_reward = calculate_reward(master1, master2, cur_done, cur_undone)
record.append([master1, master2, cur_done, cur_undone, immediate_reward])
deploy_reward.append(sum(immediate_reward))
if slot != 0:
r_grid = to_grid_rewards(immediate_reward)
targets_batch = q_estimator.compute_targets(action_mat_prev, s_grid, r_grid, gamma)
# Advantage for policy network.
advantage = q_estimator.compute_advantage(curr_state_value_prev, next_state_ids_prev,
s_grid, r_grid, gamma)
if curr_task[0][0] != -1 and curr_task[1][0] != -1:
replay.add(state_mat_prev, action_mat_prev, targets_batch, s_grid)
policy_replay.add(policy_state_prev, action_choosen_mat_prev, advantage, curr_neighbor_mask_prev)
# For updating
state_mat_prev = s_grid
action_mat_prev = valid_action_prob_mat
action_choosen_mat_prev = action_choosen_mat
curr_neighbor_mask_prev = curr_neighbor_mask
policy_state_prev = policy_state
# for computing advantage
curr_state_value_prev = curr_state_value
next_state_ids_prev = next_state_ids
global_step1 += 1
global_step2 += 1
all_rewards.append(sum(immediate_reward))
batch_reward.append(immediate_reward)
if (sum(cur_done) + sum(cur_undone)) != 0:
order_response_rates.append(float(sum(cur_done) / (sum(cur_done) + sum(cur_undone))))
else:
order_response_rates.append(0)
sum_rewards.append(float(sum(all_rewards)) / float(len(all_rewards)))
all_rewards = []
all_number = sum(achieve_num) + sum(fail_num)
throughput_list.append(sum(achieve_num) / float(all_number))
print('throughput_list_all =', throughput_list, '\ncurrent_achieve_number =', sum(achieve_num),
', current_fail_number =', sum(fail_num))
achieve_num = []
fail_num = []
episode_reward = np.sum(batch_reward[1:])
episode_rewards.append(episode_reward)
n_iter_order_response_rate = np.mean(order_response_rates[1:])
order_response_rate_episode.append(n_iter_order_response_rate)
record_all_order_response_rate.append(order_response_rates)
# update value network
for _ in np.arange(TRAIN_TIMES):
batch_s, _, batch_r, _ = replay.sample()
q_estimator.update_value(batch_s, batch_r, 1e-3, global_step1)
global_step1 += 1
# update policy network
for _ in np.arange(TRAIN_TIMES):
batch_s, batch_a, batch_r, batch_mask = policy_replay.sample()
q_estimator.update_policy(batch_s, batch_r.reshape([-1, 1]), batch_a, batch_mask, learning_rate,
global_step2)
global_step2 += 1
saver.save(sess, "./model/model.ckpt")
saver.save(sess, "./model/model_before_testing.ckpt")
tf.reset_default_graph()
time_str = str(time.time())
with open("./result/" + time_str + ".json", "w") as f:
json.dump(record, f)
return throughput_list
if __name__ == "__main__":
############ Set up according to your own needs ###########
# The parameters are set to support the operation of the program, and may not be consistent with the actual system
RUN_TIMES = 500
TASK_NUM = 5000
TRAIN_TIMES = 50
CHO_CYCLE = 1000
##############################################################
execution(RUN_TIMES, TASK_NUM, TRAIN_TIMES, CHO_CYCLE)