forked from cmantas/tiramola_v3
-
Notifications
You must be signed in to change notification settings - Fork 0
/
new_decision_module.py
684 lines (599 loc) · 35.2 KB
/
new_decision_module.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
__author__ = 'tiramola group'
import os, datetime, operator, math, random, itertools, time
import numpy as np
from lib.fuzz import fgraph, fset
from scipy.cluster.vq import kmeans2
from lib.persistance_module import env_vars
from scipy.stats import linregress
from collections import deque
from lib.tiramola_logging import get_logger
from Predictor import Predictor
class RLDecisionMaker:
def __init__(self, cluster):
#Create logger
LOG_FILENAME = 'files/logs/Coordinator.log'
self.log = get_logger('RLDecisionMaker', 'INFO', logfile=LOG_FILENAME)
self.log.info("Using 'gain' : " + env_vars['gain'] +" with threshold of "+str( env_vars["decision_threshold"]*100) + "% and interval: " + str(env_vars['decision_interval']))
self.log.info("Cluster Size from %d to %d nodes" % (env_vars['min_cluster_size'], env_vars['max_cluster_size']))
self.debug = False
if self.debug:
self.currentState = 8
else:
self.currentState = cluster.node_count()
self.cluster = cluster
self.nextState = self.currentState
self.waitForIt = env_vars['decision_interval'] / env_vars['metric_fetch_interval']
self.pending_action = None
self.decision = {"action": "PASS", "count": 0}
# The policy for getting throughput and latency when computing the reward func.
# average, centroid
self.measurementsPolicy = 'centroid'
self.prediction = env_vars['use_prediction']
self.predictor = Predictor()
# used only in simulation!!
self.countdown = 0
# A dictionary that will remember rewards and metrics in states previously visited
self.memory = {}
for i in range(env_vars["min_cluster_size"], env_vars["max_cluster_size"] + 1):
self.memory[str(i)] = {}
#self.memory[str(i)]['V'] = None # placeholder for rewards and metrics
self.memory[str(i)]['r'] = None
self.memory[str(i)]['arrayMeas'] = None
# Load any previous statics.
self.measurementsFile = env_vars["measurements_file"]
self.trainingFile = env_vars["training_file"]
self.sumMetrics = {}
# initialize measurements file
meas = open(self.measurementsFile, 'a+')
if os.stat(self.measurementsFile).st_size == 0:
# The file is empty, set the headers for each column.
meas.write('State\t\tLambda\t\tThroughput\t\tLatency\t\tCPU\t\tTime\n')
meas.close()
# load training set
meas = open(self.trainingFile, 'r+')
if os.stat(self.trainingFile).st_size != 0:
# Read the training set measurements saved in the file.
meas.next() # Skip the first line with the headers of the columns
for line in meas:
# Skip comments (used in training sets)
if not line.startswith('###'):
m = line.split('\t\t')
self.add_measurement(m)
meas.close()
def add_measurement(self, metrics, write_file=False, write_mem=True):
"""
adds the measurement to either memory or file or both
@param metrics: array The metrics to store. An array containing [state, lamdba, throughput, latency, time]
@param writeFile: boolean If set write the measurement in the txt file
:return:
"""
if self.measurementsPolicy.startswith('average'):
if not self.sumMetrics.has_key(metrics[0]):
# Save the metric with the state as key metrics = [state, inlambda, throughput, latency]
self.sumMetrics[metrics[0]] = {'inlambda': 0.0, 'throughput': 0.0, 'latency': 0.0, 'divide_by': 0}
self.sumMetrics[metrics[0]] = {'inlambda': self.sumMetrics[metrics[0]]['inlambda'] + float(metrics[1]),
'throughput': self.sumMetrics[metrics[0]]['throughput'] + float(metrics[2]),
'latency': self.sumMetrics[metrics[0]]['latency'] + float(metrics[3]),
'divide_by': self.sumMetrics[metrics[0]]['divide_by'] + 1}
if self.debug and write_file:
self.log.debug("add_measurements: won't load measurement to memory")
else:
if write_mem:
# metrics-> 0: state, 1: lambda, 2: thoughtput, 3:latency, 4:cpu, 5:time
if not self.memory.has_key(metrics[0]):
self.memory[str(metrics[0])] = {}
#self.memory[str(metrics[0])]['V'] = None # placeholder for rewards and metrics
self.memory[str(metrics[0])]['r'] = None
self.memory[str(metrics[0])]['arrayMeas'] = np.array([float(metrics[1]), float(metrics[2]),
float(metrics[3]), float(metrics[4])], ndmin=2)
elif self.memory[metrics[0]]['arrayMeas'] is None:
self.memory[metrics[0]]['arrayMeas'] = np.array([float(metrics[1]), float(metrics[2]),
float(metrics[3]), float(metrics[4])], ndmin=2)
else:
self.memory[metrics[0]]['arrayMeas'] = np.append(self.memory[metrics[0]]['arrayMeas'],
[[float(metrics[1]), float(metrics[2]),
float(metrics[3]), float(metrics[4])]], axis=0)
# but add 1 zero measurement for each state for no load cases ??? too many 0s affect centroids?
if write_file:
if write_mem:
used = "Yes"
else:
used = "No"
ms = open(self.measurementsFile, 'a')
# metrics[5] contains the time tick -- when running a simulation, it represents the current minute,
# on actual experiments, it is the current time. Used for debugging and plotting
ms.write(str(metrics[0]) + '\t\t' + str(metrics[1]) + '\t\t' + str(metrics[2]) + '\t\t' +
str(metrics[3]) + '\t\t' + str(metrics[4]) + '\t\t' + str(metrics[5]) + '\t\t'+ used+'\n')
ms.close()
# param state: string Get the average metrics (throughput, latency) for this state.
# return a dictionary with the averages
def get_averages(self, state):
averages = {}
if self.sumMetrics.has_key(state):
averages['throughput'] = float(self.sumMetrics[state]['throughput'] / self.sumMetrics[state]['divide_by'])
averages['latency'] = float(self.sumMetrics[state]['latency'] / self.sumMetrics[state]['divide_by'])
self.log.debug("GETAVERAGES Average metrics for state: " + state + " num of measurements: " + str(
self.sumMetrics[state]['divide_by']) +
" av. throughput: " + str(averages['throughput']) + " av. latency: " +
str(averages['latency']))
return averages
def doKmeans(self, state, from_inlambda, to_inlambda):
# Run kmeans for the measurements of this state and return the centroid point (throughput, latency)
ctd = {}
label = []
centroids = {}
if self.memory[state]['arrayMeas'] != None:
count_state_measurements = len(self.memory[state]['arrayMeas'])
# self.log.debug("DOKMEANS " + str(len(self.memory[state]['arrayMeas'])) +
# " measurements available for state " + state)
sliced_data = None
for j in self.memory[state]['arrayMeas']:
#self.my_logger.debug("DOKMEANS self.memory[state]['arrayMeas'][j]: "+ str(j))
# If this measurement belongs in the slice we're insterested in
if j[0] >= from_inlambda and j[0] <= to_inlambda:
#self.my_logger.debug("DOKMEANS adding measurement : "+ str(j))
# add it
if sliced_data == None:
sliced_data = np.array(j, ndmin=2)
else:
sliced_data = np.append(sliced_data, [j], axis=0)
k = 1 # number of clusters
# 1. No known lamdba values close to current lambda measurement
if sliced_data == None:
# Check if there are any known values from +-50% inlambda.
# original_inlambda = float(from_inlambda* (10/9))
# from_inlambda = 0.8 * original_inlambda
# to_inlambda = 1.2 * original_inlambda
# self.my_logger.debug("Changed lambda range to +- 20%: "+ str(from_inlambda) + " - "+ str(to_inlambda))
# for j in self.memory[state]['arrayMeas']:
# #self.my_logger.debug("DOKMEANS self.memory[state]['arrayMeas'][j]: "+ str(j))
# # If this measurement belongs in the slice we're insterested in
# if j[0] >= from_inlambda and j[0] <= to_inlambda:
# # add it
# if sliced_data == None:
# sliced_data = np.array(j, ndmin=2)
# else:
# sliced_data = np.append(sliced_data, [j], axis=0)
# #centroids, label = kmeans2(self.memory[state]['arrayMeas'], k, minit='points') # (obs, k)
# #else:
# if sliced_data == None:
self.log.debug("No known lamdba values close to current lambda measurement. Returning zeros!")
else:
# self.log.debug("DOKMEANS length of sliced_data to be fed to kmeans: " + str(len(sliced_data))
# + " (out of %d total)" % count_state_measurements)
centroids, label = kmeans2(sliced_data, k, minit='points')
pass
# initialize dictionary
num_of_meas = {}
#num_of_meas = {'0': 0, '1': 0, '2': 0, '3': 0, '4': 0}
for j in range(0, k):
num_of_meas[str(j)] = 0
if len(label) > 0:
for i in label:
num_of_meas[str(i)] += 1
max_meas_cluster = max(num_of_meas.iteritems(), key=operator.itemgetter(1))[0]
# self.my_logger.debug("DOKMEANS state: "+ state +" kmeans2 centroids: "+ str(centroids) +" label: "+
# str(num_of_meas) + " cluster with max measurements: "+ str(max_meas_cluster))
ctd['inlambda'] = centroids[int(max_meas_cluster)][0]
ctd['throughput'] = centroids[int(max_meas_cluster)][1]
ctd['latency'] = centroids[int(max_meas_cluster)][2]
ctd['cpu'] = centroids[int(max_meas_cluster)][3]
else:
#self.log.debug("DOKMEANS one of the clusters was empty and so label is None :|. Returning zeros")
ctd['inlambda'] = 0.0
ctd['throughput'] = 0.0
ctd['latency'] = 0.0
ctd['cpu'] = 0.0
#return None
else:
self.log.debug("DOKMEANS self.memory[state]['arrayMeas'] is None :|")
return ctd
def moving_average(self, iterable, n=3):
# moving_average([40, 30, 50, 46, 39, 44]) --> 40.0 42.0 45.0 43.0
# http://en.wikipedia.org/wiki/Moving_average
it = iter(iterable)
d = deque(itertools.islice(it, n - 1))
d.appendleft(0)
s = sum(d)
for elem in it:
s += elem - d.popleft()
d.append(elem)
yield s / float(n)
def predict_load(self):
# Linear Regression gia na doume to slope
stdin, stdout = os.popen2("tail -n 20 " + self.measurementsFile)
stdin.close()
lines = stdout.readlines();
stdout.close()
ten_min_l = [] # store past 10 mins lambda's
ten_min = [] # store past 10 mins ticks
for line in lines:
m = line.split('\t\t') # state, lambda, throughput, latency, cpu, time tick
ten_min_l.append(float(m[1]))
ten_min.append(float(m[5]))
# run running average on the 10 mins lambda measurements
n = 5
run_avg_gen = self.moving_average(ten_min_l, n)
run_avg = []
for r in run_avg_gen:
run_avg.append(float(r))
ten_min_ra = ten_min[2:18] # np.arange(i-8, i-2, 1)
# linear regression on the running average
#(slope, intercept, r_value, p_value, stderr) = linregress(ten_min, ten_min_l)
(slope, intercept, r_value, p_value, stderr) = linregress(ten_min_ra, run_avg)
# fit the running average in a polynomial
coeff = np.polyfit(ten_min, ten_min_l, deg=2)
self.log.debug("Slope (a): " + str(slope) + " Intercept(b): " + str(intercept))
self.log.debug("Polynom coefficients: " + str(coeff))
#self.my_logger.debug("next 10 min prediction "+str(float(slope * (p + 10) + intercept + stderr)))
predicted_l = float(slope * (ten_min[19] + 10) + intercept + stderr) # lambda in 10 mins from now
#predicted_l = np.polyval(coeff, (ten_min[9] + 10)) # lambda in 10 mins from now
if slope > 0:
#if predicted_l > allmetrics['inlambda'] :
dif = 6000 + 10 * int(slope)
#dif = 6000 + 0.2 * int(predicted_l - allmetrics['inlambda'])
self.log.debug("Positive slope: " + str(slope) + " dif: " + str(dif)
+ ", the load is increasing. Moving the lambda slice considered 3K up")
else:
dif = -6000 + 10 * int(slope)
#dif = -6000 + 0.2 * int(predicted_l - allmetrics['inlambda'])
self.log.debug("Negative slope " + str(slope) + " dif: " + str(dif)
+ ", the load is decreasing. Moving the lambda slice considered 3K down")
#dif = ((predicted_l - allmetrics['inlambda'])/ allmetrics['inlambda']) * 0.1 * 6000#* allmetrics['inlambda']
#dif = int((predicted_l / allmetrics['inlambda']) * 6000)
return predicted_l
def publish_to_local_ganglia(self, allmetrics):
"""
Publishes monitoring data to local ganglia agent
:param allmetrics:
:return:
"""
self.log.debug( "TAKEDECISION allmetrics: " + str(allmetrics))
#Publish measurements to ganglia
try:
os.system("gmetric -n ycsb_inlambda -v " + str(
allmetrics['inlambda']) + " -d 15 -t float -u 'reqs/sec' -S " + str(
self.monitoring_endpoint) + ":[DEBUG] hostname")
os.system("gmetric -n ycsb_throughput -v " + str(
allmetrics['throughput']) + " -d 15 -t float -u 'reqs/sec' -S " + str(
self.monitoring_endpoint) + ":[DEBUG] hostname")
os.system(
"gmetric -n ycsb_latency -v " + str(allmetrics['latency']) + " -d 15 -t float -u ms -S " + str(
self.monitoring_endpoint) + ":[DEBUG] hostname")
except:
pass
def handle_metrics(self, client_metrics, server_metrics):
# read metrics
allmetrics = {'inlambda': 0, 'throughput': 0, 'latency': 0, 'cpu': 0}
if not self.debug:
## Aggreggation of YCSB client metrics
clients = 0
servers = 0
# We used to collect server cpu too, do we need it?
#self.log.debug("TAKEDECISION state: %d, pending action: %s. Collecting metrics" % (self.currentState, str(self.pending_action)))
for host in client_metrics.keys():
metric = client_metrics[host]
if isinstance(metric, dict):
for key in metric.keys():
if key.startswith('ycsb_TARGET'):
allmetrics['inlambda'] += float(metric[key])
elif key.startswith('ycsb_THROUGHPUT'):
allmetrics['throughput'] += float(metric[key])
elif key.startswith('ycsb_READ') or key.startswith('ycsb_UPDATE') or key.startswith(
'ycsb_RMW') or key.startswith('ycsb_INSERT'):
allmetrics['latency'] += float(metric[key])
clients += 1
for host in server_metrics.keys():
metric = server_metrics[host]
if isinstance(metric, dict):
#check if host in active cluster hosts
if not host in self.cluster.get_hosts().keys():
continue
servers += 1
for key in metric.keys():
if key.startswith('cpu_idle'):
allmetrics['cpu'] += float(metric[key])
try:
allmetrics['latency'] = allmetrics['latency'] / clients
except:
allmetrics['latency'] = 0
try:
allmetrics['cpu'] = (allmetrics['cpu'] / servers) # average node cpu usage
except:
allmetrics['cpu'] = 0
else:
self.log.info("Running in DEBUG mode, no metrics retrieved!")
return allmetrics
# a log-related variable
pending_action_logged = False
def take_decision(self, client_metrics, server_metrics):
'''
this method reads allmetrics object created by Monitoring.py and decides whether a change
of the number of participating
virtual nodes is due.
'''
# update prediction current minute counter
self.predictor.tick_tock()
if client_metrics is None or server_metrics is None: return
# first parse all metrics
allmetrics = self.handle_metrics(client_metrics, server_metrics)
#self.publish_to_local_ganglia(allmetrics)
pending_action = not (self.pending_action is None) # true if there is no pending action
# 1. Save the current metrics to file and in memory only if there is no pending action.
self.add_measurement([str(self.currentState), allmetrics['inlambda'], allmetrics['throughput'],
allmetrics['latency'], allmetrics['cpu'],
datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")],
write_file=True, write_mem=((not pending_action) and bool(env_vars['update_metrics'])))
# if there is a pending action, don't take a decision
if pending_action:
global pending_action_logged
if not pending_action_logged:
self.log.debug("Last action " + self.pending_action + " hasn't finished yet, see you later!")
pending_action_logged = True
if self.debug:
if self.countdown == 0:
self.log.debug("Running a simulation, set state from " + str(self.currentState) + " to " +
str(self.nextState))
self.currentState = self.nextState
self.pending_action = None
else:
self.countdown -= 1
self.log.debug("Reducing countdown to " + str(self.countdown))
# skip decision
self.decision["action"] = "PASS"
self.decision["count"] = 0
return self.decision
pending_action_logged = False
# manage the interval counter (waitForIt)
if self.waitForIt == 0:
self.waitForIt = env_vars['decision_interval'] / env_vars['metric_fetch_interval']
else:
if self.waitForIt == env_vars['decision_interval'] / env_vars['metric_fetch_interval']:
self.log.debug("New decision in " + str(float(self.waitForIt*env_vars['metric_fetch_interval'])/60) +
" mins, see you later!")
self.waitForIt -= 1
self.decision["action"] = "PASS"
self.decision["count"] = 0
return self.decision
# Select values close to the current throughtput, define tha lambda range we're interested in -+ 5%
slice_range=75
from_inlambda = allmetrics['inlambda'] - slice_range
to_inlambda = allmetrics['inlambda'] + slice_range
if self.prediction:
predicted_l = self.predictor.poly_regression()
if predicted_l > 0:
# there are enough data to make a prediction, if not use the actual lambda
self.log.debug(
"Predicted: " + str(predicted_l) + " lambda :" + str(allmetrics['inlambda']))
from_inlambda = predicted_l - slice_range
to_inlambda = predicted_l + slice_range
self.log.debug("TAKEDECISION state %d lambda range: %d - %d" % (self.currentState, from_inlambda, to_inlambda))
# too low to care, the initial num of nodes can answer 1000 req/sec,
# so consider it as 0 1000 * len(cluster.size)!!
if 0.0 < to_inlambda < 1000:
from_inlambda = 0.0
self.log.debug("TAKEDECISION state %d current lambda %d changed lambda range to: %d - %d"
% (self.currentState, allmetrics['inlambda'], from_inlambda, to_inlambda))
# The subgraph we are interested in. It contains only the allowed transitions from the current state.
from_node = max(int(env_vars["min_cluster_size"]), (self.currentState - env_vars["rem_nodes"]))
to_node = min(self.currentState + int(env_vars["add_nodes"]), int(env_vars["max_cluster_size"]))
#self.my_logger.debug("TAKEDECISION creating graph from node: "+ str(from_node) +" to node "+ str(to_node))
#inject the current number of nodes
allmetrics['current_nodes'] = self.currentState
states = fset.FuzzySet()
# Calculate rewards using the values in memory if any, or defaults
for i in range(from_node, to_node + 1):
# se periptwsi pou den exeis 3anadei to state upologizei poso tha ithele na einai to throughput
# allmetrics['max_throughput'] = float(i) * float(self.utils.serv_throughput)
allmetrics['num_nodes'] = i
met = {}
if self.measurementsPolicy.startswith('average'):
met = self.getAverages(str(i))
elif self.measurementsPolicy.startswith('centroid'):
met = self.doKmeans(str(i), from_inlambda, to_inlambda)
#format met output
out_met = {k: int(v) for k,v in met.iteritems()}
self.log.debug("TAKEDECISION state: " + str(i) + " met: " + str(out_met))
if met != None and len(met) > 0:
# Been in this state before, use the measurements
allmetrics['inlambda'] = met['inlambda']
allmetrics['throughput'] = met['throughput']
allmetrics['latency'] = met['latency']
allmetrics['cpu'] = met['cpu']
#self.my_logger.debug("TAKEDECISION adding visited state "+ str(i) +" with gain "+ str(self.memory[str(i)]['r']))
#else:
# No clue for this state use current measurements...
#self.my_logger.debug("TAKEDECISION unknown state "+ str(i) +" with gain "+ str(self.memory[str(i)]['r']))
self.memory[str(i)]['r'] = eval(env_vars["gain"], allmetrics)
# if self.currentState != i:
# self.my_logger.debug(
# "TAKEDECISION adding state " + str(i) + " with gain " + str(self.memory[str(i)]['r']))
states.add(fset.FuzzyElement(str(i), self.memory[str(i)]['r']))
# For the current state, use current measurement
# if self.currentState == i:
# if not self.debug:
# cur_gain = eval(env_vars["gain"], allmetrics)
# # for debugging purposes I compare the current reward with the one computed using the training set
# self.log.debug("TAKEDECISION state %d current reward: %d training set reward: %d"
# % (self.currentState, cur_gain, self.memory[str(i)]['r']))
# self.memory[str(i)]['r'] = cur_gain
# #self.log.debug("TAKEDECISION adding current state " + str(i) + " with gain " + str(cur_gain))
# else:
# cur_gain = (self.memory[str(i)]['r'])
# self.log.debug("TAKEDECISION state %d current state training set reward: %d"
# % (self.currentState, cur_gain))
#
# states.add(fset.FuzzyElement(str(i), cur_gain))
# Create the transition graph
v = []
for i in states.keys():
v.append(i)
v = set(v)
stategraph = fgraph.FuzzyGraph(viter=v, directed=True)
for j in range(from_node, to_node + 1):
if j != self.currentState:
# Connect nodes with allowed transitions from the current node.connect(tail, head, mu) head--mu-->tail
stategraph.connect(str(j), str(self.currentState), eval(env_vars["trans_cost"], allmetrics))
#self.my_logger.debug(
# "TAKEDECISION connecting state " + str(self.currentState) + " with state " + str(j))
# Connect nodes with allowed transitions from node j.
#for k in range(max(int(env_vars["min_cluster_size"]), j - int(env_vars["rem_nodes"])),
# min(j + int(env_vars["add_nodes"]), int(env_vars["max_cluster_size"])+1)):
# if k != j:
# self.my_logger.debug("TAKEDECISION connecting state "+ str(j) +" with state "+ str(k))
# stategraph.connect(str(k), str(j), eval(env_vars["trans_cost"], allmetrics))
#Calculate the V matrix for available transitions
V = {}
for s in range(from_node, to_node + 1):
# Get allowed transitions from this state.
if self.memory[str(s)]['r'] != None:
# For each state s, we need to calculate the transitions allowed.
#allowed_transitions = stategraph.edges(head=str(s))
#Vs = []
# for t in allowed_transitions:
# t[0] is the tail state of the edge (the next state)
# No V from last run
#if self.memory[t[0]]['V'] == None:
# self.memory[t[0]]['V'] = self.memory[t[0]]['r']
# Vs.append(self.memory[t[0]]['r'])
# self.my_logger.debug("TAKEDECISION tail state: "+ t[0] +" head state: "+
# t[1] +" V("+t[0]+") = "+ str(self.memory[t[0]]['V']))
# self.my_logger.debug("TAKEDECISION transition cost from state:"+ str(t[1]) +" to state: "+ str(t[0]) +
# " is "+ str(stategraph.mu(t[1],t[0])))
# The original algo uses previous values of max reward (+ gamma * previous max), we don't
# if len(Vs) > 0:
# V[s] = self.memory[str(s)]['r'] + float(self.utils.gamma) * max(Vs)
# else:
# V[s] = self.memory[str(s)]['r']
V[s] = self.memory[str(s)]['r']
self.log.debug("TAKEDECISION Vs="+str(V))
# Find the max V (the min state with the max value)
max_gain = max(V.values())
max_set = [key for key in V if V[key] == max_gain]
self.log.debug("max set: "+str(max_set))
self.nextState = min(max_set)
self.log.debug("max(V): %d (GAIN=%d)" % (self.nextState, V[self.nextState]))
#self.my_logger.debug("TAKEDECISION next state: "+ str(self.nextState))
# Remember the V values calculated ???
#for i in V.keys():
# self.memory[str(i)]['V'] = V[i]
# self.my_logger.debug("TAKEDECISION V("+ str(i) +") = "+ str(V[i]))
# vis = fuzz.visualization.VisManager.create_backend(stategraph)
# (vis_format, data) = vis.visualize()
#
# with open("%s.%s" % ("states", vis_format), "wb") as fp:
# fp.write(data)
# fp.flush()
# fp.close()
if self.nextState != self.currentState:
self.log.debug("Decided to change state to_next: " + str(self.nextState) + " from_curr: " + str(self.currentState))
# You've chosen to change state, that means that nextState has a greater reward, therefore d is always > 0
current_reward = self.memory[str(self.currentState)]['r']
d = self.memory[str(self.nextState)]['r'] - current_reward
self.log.debug( "Difference is " + str(d) + " abs thres="+str(env_vars['decision_abs_threshold'])+" gte:"+str(float(d) < env_vars['decision_abs_threshold']))
if (current_reward != 0 and (abs(float(d) / current_reward) < env_vars['decision_threshold']))\
or float(d) < env_vars['decision_abs_threshold']:
#false alarm, stay where you are
self.nextState = self.currentState
# skip decision
self.decision["action"] = "PASS"
self.decision["count"] = 0
self.log.debug("ups changed my mind...staying at state: " + str(self.currentState) +
" cause the gain difference is: " + str(abs(d)) +
" which is less than %d%% of the current reward, it's actually %f%%" % (int(100*env_vars['decision_threshold']) ,abs(float(d)*100) / (float(current_reward)+0.001)))
else:
self.log.debug("Difference "+ str(d) + " is greater than threshold ("+str(env_vars['decision_threshold'])+"). Keeping decision")
# If the reward is the same with the state you're in, don't move
# elif (d == 0):
# #false alarm, stay where you are
# self.nextState = self.currentState
# # skip decision
# self.decision["action"] = "PASS"
# self.decision["count"] = 0
# self.log.debug("ups changed my mind...staying at state: " + str(self.currentState) +
# " cause the gain difference is: " + str(abs(d)) +
# " which is less than 10% of the current reward "
# + str(self.memory[str(self.currentState)]['r']))
if self.nextState > self.currentState:
self.decision["action"] = "ADD"
elif self.nextState < self.currentState:
self.decision["action"] = "REMOVE"
self.decision["count"] = abs(int(self.currentState) - int(self.nextState))
#self.log.debug("TAKEDECISION: action " + self.decision["action"] + " " + str(self.decision["count"]) +
# " nodes.")
## Don't perform the action if we're debugging/simulating!!!
if self.debug:
if self.pending_action is None and not self.decision["action"].startswith("PASS"):
self.pending_action = self.decision['action']
self.countdown = 2 * self.decision['count'] * 60 / env_vars['metric_fetch_interval']
#self.currentState = str(self.nextState)
self.log.debug("TAKEDECISION simulation, action will finish in: " + str(self.countdown) + " mins")
else:
self.log.debug("TAKEDECISION Waiting for action to finish: " + str(self.pending_action))
return self.decision
def simulate(self):
self.log.debug("START SIMULATION!!")
## creates a sin load simulated for an hour
# for i in range(0, 3600, 10):
#for i in range(0, 14400, 60): # 4 hours
for i in range(0, 900, 1):
cpu = max(5, 60 * abs(math.sin(0.05 * math.radians(i))) - int(self.currentState))
# lamdba is the query arrival rate, throughput is the processed queries
#l = 60000 + 40000 * math.sin(0.01 * i) + random.uniform(-4000, 4000)
#l = 50000 * math.sin(60 * math.radians(i)/40) + 65000 + random.uniform(-8000, 8000)
#l = 40000 * math.sin(60 * math.radians(i)/50) + 45000 + random.uniform(-4000, 4000)
#l = 30000 * math.sin(0.02 * i) + 55000 + random.uniform(-4000, 4000)
l = 60000 * math.sin(0.04 * i) + 75000 + random.uniform(-6000, 6000)
# first 10 mins
# if i < 1200:
# l = 20000
# elif i < 2400:
# l = 40000
# elif i < 4400:
# l = 60000
# elif i < 6000:
# l = 40000
# elif i < 7200:
# l = 20000
maxThroughput = (float(self.currentState) * float(env_vars["serv_throughput"]))
# latency = 200 # msec
# if (l > maxThroughput):
# latency += (l-maxThroughput)/10 # +100msec for every 1000 reqs queued
#throughput = min(maxThroughput, l)# max throughput for the current cluster
throughput = l #(+/- e ??)
latency = 0.0000004 * l ** 2 + 200 # msec...
if l > maxThroughput:
throughput = maxThroughput - 0.01 * l
latency = 0.00001 * (l - maxThroughput) ** 2 + (0.0000004 * maxThroughput ** 2 + 200) # msec... ?
values = {'latency': latency, 'cpu': cpu, 'inlambda': l, 'throughput': throughput,
'num_nodes': self.currentState}
self.log.debug(
"SIMULATE i: " + str(i) + " state: " + str(self.currentState) + " values:" + str(values)
+ " maxThroughput: " + str(maxThroughput))
#nomizw de xreiazetai giati ginetai kai take_decision kai se debug mode
#self.addMeasurement([self.currentState, str(l), str(throughput), str(latency), str(i)], True)
#if self.pending_action[len(self.pending_action)-1] == "done" :
self.take_decision(values)
time.sleep(1)
return
def simulate_training_set(self):
# run state 12 lambdas
self.log.debug("START SIMULATION!!")
self.debug = True
load = []
for k in range(9, 19):
for j in self.memory[str(k)]['arrayMeas']:
load.append(j[0])
#for i in range(0, 120, 1): # paizei? 1 wra ana miso lepto
for i in range(0, 240*12, 1):
l = load[i]
# throughput = (800 * self.currentState)
# if l < (800 * self.currentState):
# throughput = l
values = {'inlambda': l, 'num_nodes': self.currentState}
self.log.debug(
"SIMULATE i: " + str(i) + " state: " + str(self.currentState) + " values:" + str(values))
self.take_decision(values)
if __name__ == '__main__':
fsm = RLDecisionMaker("localhost")
fsm.simulate_training_set()