forked from SymbioticLab/Tiresias
-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrun_sim.py
executable file
·1757 lines (1527 loc) · 72.1 KB
/
run_sim.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
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import csv
import sys
import time
import math
import copy
import os
from core import util
from core import flags
from core.jobs import base_factory
import log_manager as lm
from core.scheduling import schedule as sche
from infra import infrastructure as cluster
import logging
import datetime
# parse input arguments
flags.DEFINE_string('trace_file', 'tf_job.csv',
'''Provide TF job trace file (*.csv, *.txt).
*.csv file, use \',\' as delimiter; *.txt file, user \' \' as deliminter.
Default file is tf_job.csv ''')
flags.DEFINE_string('log_path', 'result-' + time.strftime("%Y%m%d-%H-%M-%S", time.localtime()),
'''Simulation output folder, including cluster/node/gpu usage trace, pending job_queue info.
Default folder is result-[time]''')
flags.DEFINE_string('scheme', 'yarn',
'''
Job placement scheme:
0.count, just resource counting, without assignment (which gpu, which cpu)
1.yarn, ms yarn
2.random
3.crandom (consolidate + random)
4.greedy
5.balance
6.cbalance (consolidate + balance)
Default is yarn''')
flags.DEFINE_string('schedule', 'fifo',
'''
Job schedule scheme:
1.fifo
2.fjf, fit job first( in fifo order)
3.sjf, smallest job first
4.lpjf, longest pending job first
5.shortest, shortest-remaining-time job first
6.shortest-gpu, shortest-remaining-gputime job first
7.dlas, discretized las
8.dlas-gpu, dlas using gpu time
Default is fifo''')
flags.DEFINE_boolean('pack', False, '''Enable packing for gpus job.''')
flags.DEFINE_integer('num_switch', 1, '''Part of cluster spec: the number of switches in this cluster, default is 1''')
flags.DEFINE_integer('num_node_p_switch', 32,
'''Part of cluster spec: the number of nodes under a single switch, default is 32''')
flags.DEFINE_boolean('enable_network_costs', False,
'''Enable network costs when communicating over more nodes, default is False''')
flags.DEFINE_boolean('enable_migration', False,
'''Enable preempt and migrate during scheduling''')
flags.DEFINE_integer('bandwidth', 1250,
'''
Bandwidth per rack in MBps, default 10Gbps (GigaBit)
http://arxiv.org/abs/1805.07891
modern cloud is about 10Gbps
''')
flags.DEFINE_float('internode_latency', 0.015,
'''
internode, within the same rack, should there be latency
in seconds
''')
flags.DEFINE_integer('gpu_memory_capacity', 32,
'''
Specify GPU memory capacity for the gpu,
will be used in packing scheme .
''')
flags.DEFINE_integer('num_queue', 1, '''The number of queues within the job manager''')
flags.DEFINE_integer('num_buffer', 5, '''The number of jobs to be considered within the job manager''')
flags.DEFINE_integer('num_gpu_p_node', 8,
'''Part of cluster spec: the number of gpus on each node, default is 8''')
flags.DEFINE_integer('num_cpu_p_node', 128,
'''Part of cluster spec: the number of cpus on each node, default is 64''')
flags.DEFINE_integer('mem_p_node', 512,
'''Part of cluster spec: memory capacity on each node, default is 128''')
flags.DEFINE_string('cluster_spec', None,
'''Part of cluster spec: cluster infra spec file,
this file will overwrite the specs from num_switch, num_node_p_switch, and num_gpu_p_node
Spec format:
num_switch,num_node_p_switch,num_gpu_p_node
int,int,int''')
flags.DEFINE_boolean('print', False,
'''Enable print out information, default is False''')
flags.DEFINE_boolean('flush_stdout', True,
'''Flush stdout, default is True''')
flags.DEFINE_version('0.1')
FLAGS = flags.FLAGS
'''
Gandiva scheduler assumption
'''
def gandiva_sim_jobs(gputime=False, solve_starvation=0):
'''
new jobs are added to the end of the ending queue
but any fit job should be executed in fifo order
'''
cur_time = JOBS.job_events[0]['time']
node_release = False
time_diff = 0
while (len(JOBS.job_events) + len(JOBS.pending_jobs) + len(JOBS.running_jobs))> 0:
# if len(JOBS.job_events) == 0:
# break
CLUSTER.gandiva_node_set_adjust(cur_time, JOBS, LOG)
print("%d-%d, %d, %d " % (cur_time, len(JOBS.job_events), len(JOBS.pending_jobs), len(JOBS.running_jobs)))
#update job progress for end_jobs
node_release = CLUSTER.time_slicing_execute(cur_time, JOBS, LOG, time_diff)
#get new start job
event = util.search_dict_list(JOBS.job_events, 'time', cur_time)
if event != None:
#for new-start jobs, try to start
for s_job in event['start_jobs']:
ret = scheduler.try_get_job_res(CLUSTER, JOBS, s_job)
if ret == False:
JOBS.move_to_pending(s_job)
else:
s_job['start_time'] = cur_time
JOBS.running_jobs.append(s_job)
util.print_fn('----job[%d] starts' % s_job['job_idx'])
#remove time_event
JOBS.job_events.remove(event)
if node_release:
for p_job in JOBS.pending_jobs:
ret = scheduler.try_get_job_res(CLUSTER, JOBS, p_job)
if ret == True:
JOBS.remove_from_pending(p_job, cur_time)
p_job['start_time'] = cur_time
JOBS.running_jobs.append(p_job)
util.print_fn('----job[%d] starts from pending' % p_job['job_idx'])
node_release = False
if len(JOBS.job_events) <= 0:
cur_time = cur_time + 10
time_diff = 10
else:
JOBS.job_events.sort(key = lambda e:e.__getitem__('time'))
next_e_time = JOBS.job_events[0]['time']
if (cur_time + 10) > next_e_time:
time_diff = int(next_e_time - cur_time)
cur_time = next_e_time
else:
cur_time = cur_time + 10
time_diff = 10
# cur_time = cur_time + 1
def smallest_first_sim_jobs(gputime=False):
'''
new jobs are added to the end of the ending queue
but in the queue, shortest (gpu) job first be served, until no resource
'''
end_events = list()
while (len(JOBS.job_events) + len(JOBS.runnable_jobs))> 0:
if (len(JOBS.job_events) + len(end_events)) == 0:
util.print_fn("This cluster is not large enough to run the job")
break
#decide which is the next event: start or end ?
start_time = sys.maxsize
if len(JOBS.job_events) > 0:
start_event = JOBS.job_events[0]
start_time = start_event['time']
end_time = sys.maxsize
if len(end_events) > 0:
end_event = end_events[0]
end_time = end_event['time']
if end_time < start_time:
event_time = end_time
event = end_events[0]
elif end_time > start_time:
event_time = start_time
# print("start-time %d, end_time %d" % (start_time, end_time))
event = JOBS.job_events.pop(0)
elif end_time == start_time and end_time != sys.maxsize:
event_time = start_time
event = JOBS.job_events.pop(0)
event['end_jobs'] = end_events[0]['end_jobs']
assert event_time == event['time']
#for ending jobs, release gpu
if 'end_jobs' in event:
for e_job in event['end_jobs']:
#job completes
CLUSTER.release_job_res(e_job)
# CLUSTER.release_gpus(e_job)
LOG.job_complete(e_job, event_time)
JOBS.runnable_jobs.remove(e_job)
#for new-start jobs, add to runnable
if 'start_jobs' in event:
for s_job in event['start_jobs']:
#add into runnable list with pending status
JOBS.move_to_runnable(s_job)
s_job['remaining_time'] = s_job['duration']
s_job['remaining_gputime'] = int(s_job['remaining_time'] * s_job['num_gpu'])
util.print_fn('---- job[%d] is added' % s_job['job_idx'])
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['total_executed_time'] = int(rjob['total_executed_time'] + tmp)
rjob['last_check_time'] = event_time
rjob['remaining_time'] = int(rjob['duration'] - rjob['total_executed_time'])
if gputime:
rjob['remaining_gputime'] = int(rjob['remaining_time'] * rjob['num_gpu'])
elif 'PENDING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['pending_time'] = int(rjob['pending_time'] + tmp)
rjob['last_check_time'] = event_time
elif 'END' == rjob['status']: #almost impossible
JOBS.runnable_jobs.remove(rjob)
pass
#sort jobs with shortest first
# if gputime:
# JOBS.runnable_jobs.sort(key = lambda e:e.__getitem__('remaining_gputime'))
# else:
# JOBS.runnable_jobs.sort(key = lambda e:e.__getitem__('remaining_time'))
JOBS.runnable_jobs.sort(key = lambda e:e.__getitem__('num_gpu'))
run_jobs = list()
preempt_jobs = list()
#scan / execute jobs one by one
CLUSTER.empty_infra()
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
if 'placements' in rjob:
del rjob['placements'][:]
ret = scheduler.try_get_job_res(CLUSTER, JOBS, rjob)
if True == ret:
# rjob['status'] = 'RUNNING'
# if 0 == rjob['start_time'] and 0 != rjob['submit_time']:
# rjob['start_time'] = event_time
if sys.maxsize == rjob['start_time']:
rjob['start_time'] = event_time
if rjob['status'] == 'PENDING':
run_jobs.append(rjob)
else:
# rjob['status'] = 'PENDING'
if rjob['status'] == 'RUNNING':
preempt_jobs.append(rjob)
continue
for job in preempt_jobs:
job['status'] = 'PENDING'
job['preempt'] = int(job['preempt'] + 1)
for job in run_jobs:
job['status'] = 'RUNNING'
job['resume'] = int(job['resume'] + 1)
# get the next end_event
del end_events[:]
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
end_time = int(event_time + rjob['remaining_time'])
tmp_dict = util.search_dict_list(end_events, 'time', end_time)
if tmp_dict == None:
#not found, add the time into to job_events
tmp_dict = dict()
tmp_dict['time'] = end_time
tmp_dict['end_jobs'] = list()
tmp_dict['end_jobs'].append(rjob)
end_events.append(tmp_dict)
else:
tmp_dict['end_jobs'].append(rjob)
end_events.sort(key = lambda e:e.__getitem__('time'))
LOG.checkpoint(JOBS, event_time)
def cal_shortest_expected_remaining(job_data, a):
data = job_data['data']
idx = next(x[0] for x in enumerate(data) if x[1] > a)
if idx == (job_data['num'] - 1):
return data[idx]
num = job_data['num'] - 1 - idx
return round(sum(data[idx: (job_data['num'] - 1)]) * 1.0 / num, 2)
def shortest_first_sim_jobs(gputime=False):
'''
new jobs are added to the end of the ending queue
but in the queue, shortest (gpu) job first be served, until no resource
'''
end_events = list()
while (len(JOBS.job_events) + len(JOBS.runnable_jobs)) > 0:
if (len(JOBS.job_events) + len(end_events)) == 0:
util.print_fn("This cluster is not large enough to run the job")
break
# decide which is the next event: start or end ?
start_time = sys.maxsize
if len(JOBS.job_events) > 0:
start_event = JOBS.job_events[0]
start_time = start_event['time']
end_time = sys.maxsize
if len(end_events) > 0:
end_event = end_events[0]
end_time = end_event['time']
if end_time < start_time:
event_time = end_time
event = end_events[0]
elif end_time > start_time:
event_time = start_time
# print("start-time %d, end_time %d" % (start_time, end_time))
event = JOBS.job_events.pop(0)
elif end_time == start_time and end_time != sys.maxsize:
event_time = start_time
event = JOBS.job_events.pop(0)
event['end_jobs'] = end_events[0]['end_jobs']
assert event_time == event['time']
# for ending jobs, release gpu
if 'end_jobs' in event:
for e_job in event['end_jobs']:
# job completes
CLUSTER.release_job_res(e_job)
# CLUSTER.release_gpus(e_job)
LOG.job_complete(e_job, event_time)
JOBS.runnable_jobs.remove(e_job)
# for new-start jobs, add to runnable
if 'start_jobs' in event:
for s_job in event['start_jobs']:
# add into runnable list with pending status
JOBS.move_to_runnable(s_job)
if FLAGS.schedule == 'shortest-expected':
s_job['remaining_expected'] = cal_shortest_expected_remaining(JOBS.job_dist_data, 0)
s_job['remaining_time'] = s_job['duration']
s_job['remaining_gputime'] = int(s_job['remaining_time'] * s_job['num_gpu'])
util.print_fn('---- job[%d] is added' % s_job['job_idx'])
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['total_executed_time'] = int(rjob['total_executed_time'] + tmp)
rjob['last_check_time'] = event_time
rjob['remaining_time'] = int(rjob['duration'] - rjob['total_executed_time'])
if FLAGS.schedule == 'shortest-expected':
s_job['remaining_expected'] = cal_shortest_expected_remaining(JOBS.job_dist_data, rjob['total_executed_time'])
if gputime:
rjob['remaining_gputime'] = int(rjob['remaining_time'] * rjob['num_gpu'])
elif 'PENDING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['pending_time'] = int(rjob['pending_time'] + tmp)
rjob['last_check_time'] = event_time
elif 'END' == rjob['status']: #almost impossible
JOBS.runnable_jobs.remove(rjob)
pass
#sort jobs with shortest first
if FLAGS.schedule == 'shortest-expected':
JOBS.runnable_jobs.sort(key = lambda e:e.__getitem__('remaining_expected'))
else:
if gputime:
JOBS.runnable_jobs.sort(key = lambda e:e.__getitem__('remaining_gputime'))
else:
JOBS.runnable_jobs.sort(key = lambda e:e.__getitem__('remaining_time'))
run_jobs = list()
preempt_jobs = list()
#scan / execute jobs one by one
CLUSTER.empty_infra()
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
if 'placements' in rjob:
del rjob['placements'][:]
ret = scheduler.try_get_job_res(CLUSTER, JOBS, rjob)
if True == ret:
# rjob['status'] = 'RUNNING'
# if 0 == rjob['start_time'] and 0 != rjob['submit_time']:
# rjob['start_time'] = event_time
if sys.maxsize == rjob['start_time']:
rjob['start_time'] = event_time
if rjob['status'] == 'PENDING':
run_jobs.append(rjob)
else:
# rjob['status'] = 'PENDING'
if rjob['status'] == 'RUNNING':
preempt_jobs.append(rjob)
continue
for job in preempt_jobs:
job['status'] = 'PENDING'
job['preempt'] = int(job['preempt'] + 1)
for job in run_jobs:
job['status'] = 'RUNNING'
job['resume'] = int(job['resume'] + 1)
# get the next end_event
del end_events[:]
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
end_time = int(event_time + rjob['remaining_time'])
tmp_dict = util.search_dict_list(end_events, 'time', end_time)
if tmp_dict == None:
#not found, add the time into to job_events
tmp_dict = dict()
tmp_dict['time'] = end_time
tmp_dict['end_jobs'] = list()
tmp_dict['end_jobs'].append(rjob)
end_events.append(tmp_dict)
else:
tmp_dict['end_jobs'].append(rjob)
end_events.sort(key = lambda e:e.__getitem__('time'))
LOG.checkpoint(JOBS, event_time)
def multi_dlas_sim_jobs(gputime=False, solve_starvation=0):
'''
1. diffrent #GPU-job(s) should have different DLAS threshold
2. the resource under each job category has its pattern, and should be determined dynamically
TODO: each job category has separate dlas-scheduler queues;
'''
end_events = list()
next_job_jump = sys.maxsize
#interval for update cluster resource allocation
reserve_interval = 600
next_reserve = reserve_interval
# while (len(JOBS.job_events) + len(JOBS.runnable_jobs))> 0:
while (len(JOBS.job_events) + int(sum(len(k.runnable_jobs) for k in JOBS.gpu_job.values())))> 0:
if (len(JOBS.job_events) + len(end_events)) == 0:
util.print_fn("This cluster is not large enough to run the job")
break
#decide which is the next event: start or end ?
start_event = None
start_time = sys.maxsize
if len(JOBS.job_events) > 0:
start_event = JOBS.job_events[0]
start_time = start_event['time']
end_event = None
end_time = sys.maxszie
if len(end_events) > 0:
end_event = end_events[0]
end_time = end_event['time']
event_time = sys.maxsize
event = dict()
event['time'] = sys.maxsize
if end_time < start_time:
event_time = end_time
event = end_event
elif end_time > start_time:
event_time = start_time
# event = JOBS.job_events.pop(0)
event = start_event
elif end_time == start_time and end_time != sys.maxsize:
event_time = start_time
# event = JOBS.job_events.pop(0)
event = start_event
event['end_jobs'] = end_events[0]['end_jobs']
assert event_time == event['time']
#decide if job_jump first or (start/end) first
if event_time > next_job_jump:
event_time = next_job_jump
event = dict()
if event_time > next_reserve:
event_time = copy.copy(next_reserve)
event = dict()
next_reserve += reserve_interval
JOBS.reserve_gpus(CLUSTER.num_gpu)
# util.print_fn('--------------------------------- Handle event[time %d]------------------------------------' % event_time)
#for ending jobs, release gpu
if 'end_jobs' in event:
for e_job in event['end_jobs']:
'''
CLUSTER.release_job_res(e_job)
LOG.job_complete(e_job, event_time)
JOBS.runnable_jobs.remove(e_job)
JOBS.queues[e_job['q_id']].remove(e_job)
'''
JOBS.end_job(e_job)
LOG.job_complete(e_job, event_time)
util.print_fn('---- job[%d] %dG is completed' % (e_job['job_idx'], e_job['num_gpu']))
#for new jobs, append to runnable jobs with pending status
if 'start_jobs' in event:
for s_job in event['start_jobs']:
JOBS.move_to_runnable(s_job) #gpu_job operation in it
s_job['q_id'] = 0 #any new start job should be in Q0
JOBS.gpu_job[s_job['num_gpu']].queues[0].append(s_job)
util.print_fn('---- job[%d] is added' % s_job['job_idx'])
#pop start event
JOBS.job_events.pop(0)
#update end events and sort, and get the most recent one
del end_events[:]
min_end_time = sys.maxsize
tmp_end_event = dict()
next_job_jump = sys.maxsize
for num_gpu, gjob in JOBS.gpu_job.items():
if len(gjob.runnable_jobs) <= 0:
continue
print('~~~~~ working on %d-GPU %djobs~~~~~' % (num_gpu, len(gjob.runnable_jobs)), end=' ')
for rjob in gjob.runnable_jobs:
if 'RUNNING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['total_executed_time'] = int(rjob['total_executed_time'] + tmp)
rjob['executed_time'] = int(rjob['executed_time'] + tmp) # decide job priority queue
rjob['last_check_time'] = event_time
#check demotion
j_gt = 0
if gputime:
j_gt = int(rjob['executed_time'] * rjob['num_gpu'])
else:
j_gt = int(rjob['executed_time'])
cur_qid = rjob['q_id']
if cur_qid < int(gjob.num_queue - 1): #not for the last queue
if j_gt >= gjob.queue_limit[cur_qid]:
rjob['q_id'] = int(cur_qid + 1)
gjob.queues[rjob['q_id']].append(rjob)
gjob.queues[cur_qid].remove(rjob)
print("job %d demote to Q%d" % (rjob['job_idx'], rjob['q_id']))
elif 'PENDING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['last_check_time'] = event_time
rjob['pending_time'] = int(rjob['pending_time'] + tmp) #this is the total pending_time
if rjob['executed_time'] > 0: # if not started yet, job is always in Q0 and no need to push_back
rjob['last_pending_time'] = int(rjob['last_pending_time'] + tmp) #this is the total pending_time
#Q0 job no need to push_back, and must be a runned
if solve_starvation > 0 and rjob['q_id'] > 0 and rjob['total_executed_time'] > 0 and rjob['executed_time'] > 0:
if rjob['last_pending_time'] >= int(rjob['executed_time'] * solve_starvation):
rjob['executed_time'] = 0
rjob['last_pending_time'] = 0
gjob.queues[0].append(rjob)
gjob.queues[rjob['q_id']].remove(rjob)
rjob['q_id'] = 0
rjob['promote'] = int(rjob['promote'] + 1)
elif 'END' == rjob['status']: # won't happen
gjob.runnable_jobs.remove(rjob)
# util.print_fn('---- job[%d] completed' % rjob['job_idx'])
pass
''' schedule jobs in each queue '''
#empty gpu resource
gjob.empty_gpu_alloc()
# for "count" placement
run_jobs = list()
preempt_jobs = list()
for queue in gjob.queues:
for job in queue:
if gjob.alloc_free_gpus(job['num_gpu']):
# if gjob.free_gpu >= job['num_gpu']:
#should run
if job['status'] == 'PENDING':
#not running
run_jobs.append(job)
# CLUSTER.free_gpu = int(CLUSTER.free_gpu - job['num_gpu']) #done in alloc_free_gpus
else:
#should NOT run
if job['status'] == 'RUNNING':
#running
preempt_jobs.append(job)
continue
util.print_fn('%d F' % gjob.free_gpu)
for job in preempt_jobs:
job['status'] = 'PENDING'
# if job['q_id'] == 0:
# job['preempt'] = int(job['preempt'] + 1)
job['preempt'] = int(job['preempt'] + 1)
for job in run_jobs:
job['status'] = 'RUNNING'
job['resume'] = int(job['resume'] + 1)
if job['start_time'] == sys.maxsize:
job['start_time'] = event_time
#sort based on the job start time
for queue in gjob.queues:
pending_job = list()
for job in queue:
# if sys.maxsize == job['start_time'] and job['status'] == 'PENDING':
if job['status'] == 'PENDING':
pending_job.append(job)
# print(job['job_idx'])
for job in pending_job:
queue.remove(job)
queue.extend(pending_job)
#update end events and sort, and get the most recent one
for rjob in gjob.runnable_jobs:
if 'RUNNING' == rjob['status']:
remaining_time = rjob['duration'] - rjob['total_executed_time']
end_time = int(event_time + remaining_time)
if end_time < min_end_time:
tmp_end_event['time'] = end_time
tmp_end_event['end_jobs'] = list()
tmp_end_event['end_jobs'].append(rjob)
min_end_time = end_time
elif min_end_time == end_time:
tmp_end_event['end_jobs'].append(rjob)
# what's the closest queue_jump (demotion, and promotion) among all the jobs
for rjob in gjob.runnable_jobs:
if 'RUNNING' == rjob['status']:
qid = rjob['q_id']
if qid < int(gjob.num_queue - 1):
if gputime:
jump_time = int(math.ceil((gjob.queue_limit[qid] - rjob['executed_time'])/rjob['num_gpu']) + event_time)
else:
jump_time = int(gjob.queue_limit[qid] - rjob['executed_time'] + event_time)
if jump_time < next_job_jump:
next_job_jump = jump_time
elif 'PENDING' == rjob['status']: # when pending job will be push back to Q0
if solve_starvation > 0 and rjob['q_id'] > 0 and rjob['total_executed_time'] and rjob['executed_time'] > 0:
diff_time = int(rjob['executed_time'] * solve_starvation - rjob['last_pending_time'])
if diff_time > 0:
jump_time = int(diff_time + event_time)
if jump_time < next_job_jump:
next_job_jump = jump_time
if min_end_time < sys.maxsize:
end_events.append(tmp_end_event)
LOG.checkpoint_multi_dlas_gpu(event_time)
# when to update the reservation
# JOBS.reserve_gpus(CLUSTER.num_gpu)
JOBS.completion_check()
def dlas_sim_jobs(gputime=False, solve_starvation=0):
'''
Job's executed time -- priority queue
Q0:[0, 30min)
Q1:[30min,1h)
Q2:[1h, 2h)
Q3:[2h, 00)
in each queue, jobs are scheduled in fit-first with FIFO
how to avoid starvation?
TODO: 2. add move_back for avoiding starvation
'''
end_events = list()
next_job_jump = sys.maxsize
while (len(JOBS.job_events) + len(JOBS.runnable_jobs))> 0:
if (len(JOBS.job_events) + len(end_events)) == 0:
util.print_fn("This cluster is not large enough to run the job")
break
#decide which is the next event: start or end ?
start_event = None
start_time = sys.maxsize
if len(JOBS.job_events) > 0:
start_event = JOBS.job_events[0]
start_time = start_event['time']
end_event = None
end_time = sys.maxsize
if len(end_events) > 0:
end_event = end_events[0]
end_time = end_event['time']
event_time = sys.maxsize
event = dict()
event['time'] = sys.maxsize
if end_time < start_time:
event_time = end_time
event = end_event
elif end_time > start_time:
event_time = start_time
# event = JOBS.job_events.pop(0)
event = start_event
elif end_time == start_time and end_time != sys.maxsize:
event_time = start_time
# event = JOBS.job_events.pop(0)
event = start_event
event['end_jobs'] = end_events[0]['end_jobs']
assert event_time == event['time']
#decide if job_jump first or (start/end) first
if event_time > next_job_jump:
event_time = next_job_jump
event = dict()
# util.print_fn('--------------------------------- Handle event[time %d]------------------------------------' % event_time)
#for ending jobs, release gpu
if 'end_jobs' in event:
for e_job in event['end_jobs']:
CLUSTER.release_job_res(e_job)
LOG.job_complete(e_job, event_time)
# util.print_fn('---- job[%d] is completed' % e_job['job_idx'])
JOBS.runnable_jobs.remove(e_job)
JOBS.queues[e_job['q_id']].remove(e_job)
#for new jobs, append to runnable jobs with pending status
if 'start_jobs' in event:
for s_job in event['start_jobs']:
JOBS.move_to_runnable(s_job)
s_job['q_id'] = 0 #any new start job should be in Q0
JOBS.queues[0].append(s_job)
util.print_fn('---- job[%d] is added' % s_job['job_idx'])
#pop start event
JOBS.job_events.pop(0)
#update executed_time
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['total_executed_time'] = int(rjob['total_executed_time'] + tmp)
rjob['executed_time'] = int(rjob['executed_time'] + tmp) # decide job priority queue
rjob['last_check_time'] = event_time
#check demotion
j_gt = 0
if gputime:
j_gt = int(rjob['executed_time'] * rjob['num_gpu'])
else:
j_gt = int(rjob['executed_time'])
cur_qid = rjob['q_id']
if cur_qid < int(JOBS.num_queue - 1): #not for the last queue
if j_gt >= JOBS.queue_limit[cur_qid]:
rjob['q_id'] = int(cur_qid + 1)
JOBS.queues[rjob['q_id']].append(rjob)
JOBS.queues[cur_qid].remove(rjob)
print("job %d demote to Q%d" % (rjob['job_idx'], rjob['q_id']))
if FLAGS.schedule == 'dlas-gpu-gittins':
# rjob['rank'] = cal_r_gittins_index(JOBS.job_dist_data, j_gt)
rjob['rank'] = get_gittins_index(j_gt)
elif 'PENDING' == rjob['status']:
tmp = int(event_time - rjob['last_check_time'])
rjob['last_check_time'] = event_time
rjob['pending_time'] = int(rjob['pending_time'] + tmp) #this is the total pending_time
if rjob['executed_time'] > 0: # if not started yet, job is always in Q0 and no need to push_back
rjob['last_pending_time'] = int(rjob['last_pending_time'] + tmp) #this is the total pending_time
#Q0 job no need to push_back, and must be a runned
if solve_starvation > 0 and rjob['q_id'] > 0 and rjob['total_executed_time'] > 0 and rjob['executed_time'] > 0:
if rjob['last_pending_time'] >= int(rjob['executed_time'] * solve_starvation):
rjob['executed_time'] = 0
rjob['last_pending_time'] = 0
JOBS.queues[0].append(rjob)
JOBS.queues[rjob['q_id']].remove(rjob)
rjob['q_id'] = 0
rjob['promote'] = int(rjob['promote'] + 1)
if FLAGS.schedule == 'dlas-gpu-gittins':
if gputime:
j_gt = int(rjob['executed_time'] * rjob['num_gpu'])
else:
j_gt = int(rjob['executed_time'])
# rjob['rank'] = cal_r_gittins_index(JOBS.job_dist_data, j_gt)
rjob['rank'] = get_gittins_index(j_gt)
elif 'END' == rjob['status']: # won't happen
JOBS.runnable_jobs.remove(rjob)
# util.print_fn('---- job[%d] completed' % rjob['job_idx'])
pass
#push job to their new queue
# JOBS.update_priority_queues(gputime)
''' schedule jobs in each queue '''
#empty_cluster resource
CLUSTER.empty_infra()
# for "count" placement
run_jobs = list()
preempt_jobs = list()
# if FLAGS.schedule == 'dlas-gpu-gittins':
# q = JOBS.queues[0]
# q.sort(key = lambda e:(e.__getitem__('rank'), e.__getitem__('r_submit_time')), reverse=True)
for queue in JOBS.queues:
if FLAGS.schedule == 'dlas-gpu-gittins':
queue.sort(key = lambda e:(e.__getitem__('rank'), e.__getitem__('r_submit_time')), reverse=True)
for job in queue:
if CLUSTER.free_gpu >= job['num_gpu']:
#should run
if job['status'] == 'PENDING':
#not running
run_jobs.append(job)
CLUSTER.free_gpu = int(CLUSTER.free_gpu - job['num_gpu'])
else:
#should NOT run
if job['status'] == 'RUNNING':
#running
preempt_jobs.append(job)
continue
for job in preempt_jobs:
job['status'] = 'PENDING'
# if job['q_id'] == 0:
# job['preempt'] = int(job['preempt'] + 1)
job['preempt'] = int(job['preempt'] + 1)
for job in run_jobs:
job['status'] = 'RUNNING'
job['resume'] = int(job['resume'] + 1)
if job['start_time'] == sys.maxsize:
job['start_time'] = event_time
#sort based on the job start time
for queue in JOBS.queues:
#job there are many students
pending_job = list()
for job in queue:
# if sys.maxsize == job['start_time'] and job['status'] == 'PENDING':
if job['status'] == 'PENDING':
pending_job.append(job)
# print(job['job_idx'])
for job in pending_job:
queue.remove(job)
queue.extend(pending_job)
#for fit-job-first, move small running jobs in front of large pending jobs in each queue
# for queue in JOBS.queues:
# num_j = len(queue)
# #find the first pending job
# for i in range(num_j):
# if queue[i]['status'] == 'PENDING':
# break
# first_pending_idx = i
# #picking running_after_pending_jobs
# run_after_pending = list()
# for j in range(first_pending_idx, num_j):
# if queue[j]['status'] == 'RUNNING':
# run_after_pending.append(queue[j])
# #reinsert all those jobs
# for job in run_after_pending:
# queue.remove(job)
# for job in run_after_pending:
# queue.insert(i, job)
# i = int(i + 1)
# for queue in JOBS.queues:
# for job in queue:
# if 'RUNNING' == job['status']:
# if 'placements' in job:
# del job['placements'][:]
# job['status'] = 'PENDING'
# ret = try_get_job_res(job)
# if True == ret:
# job['status'] = 'RUNNING'
# if 0 == job['start_time'] and 0 != job['submit_time']:
# job['start_time'] = event_time
# else:
# job['status'] = 'PENDING'
# continue
#update end events and sort, and get the most recent one
del end_events[:]
# for rjob in JOBS.runnable_jobs:
# if 'RUNNING' == rjob['status']:
# remaining_time = rjob['duration'] - rjob['total_executed_time']
# end_time = int(event_time + remaining_time)
# tmp_dict = util.search_dict_list(end_events, 'time', end_time)
# if tmp_dict == None:
# #not found, add the time into to job_events
# tmp_dict = dict()
# tmp_dict['time'] = end_time
# tmp_dict['end_jobs'] = list()
# tmp_dict['end_jobs'].append(rjob)
# end_events.append(tmp_dict)
# else:
# tmp_dict['end_jobs'].append(rjob)
# end_events.sort(key = lambda e:e.__getitem__('time'))
min_end_time = sys.maxsize
tmp_end_event = dict()
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
remaining_time = rjob['duration'] - rjob['total_executed_time']
end_time = int(event_time + remaining_time)
if end_time < min_end_time:
tmp_end_event['time'] = end_time
tmp_end_event['end_jobs'] = list()
tmp_end_event['end_jobs'].append(rjob)
min_end_time = end_time
elif min_end_time == end_time:
tmp_end_event['end_jobs'].append(rjob)
if min_end_time < sys.maxsize:
end_events.append(tmp_end_event)
# what's the closest queue_jump (demotion, and promotion) among all the jobs
next_job_jump = sys.maxsize
for rjob in JOBS.runnable_jobs:
if 'RUNNING' == rjob['status']:
qid = rjob['q_id']
if qid < int(JOBS.num_queue - 1):
if gputime:
jump_time = int(math.ceil((JOBS.queue_limit[qid] - rjob['executed_time'])/rjob['num_gpu']) + event_time)
else:
jump_time = int(JOBS.queue_limit[qid] - rjob['executed_time'] + event_time)
if jump_time < next_job_jump:
next_job_jump = jump_time
elif 'PENDING' == rjob['status']: # when pending job will be push back to Q0
if solve_starvation > 0 and rjob['q_id'] > 0 and rjob['total_executed_time'] and rjob['executed_time'] > 0:
diff_time = int(rjob['executed_time'] * solve_starvation - rjob['last_pending_time'])
if diff_time > 0:
jump_time = int(diff_time + event_time)
if jump_time < next_job_jump:
next_job_jump = jump_time
LOG.checkpoint(JOBS, event_time)
def get_gittins_index(a):
job_info = JOBS.job_dist_data
if a > job_info['data'][-2]:
return 0
idx = next(x[0] for x in enumerate(job_info['data']) if x[1] > a)
return job_info['gittins'][idx]
def gittins_sim_jobs(job_dist_data, gputime=False, static_delta=True):
'''
gittins index
'''
solve_starvation = 0
end_events = list()
next_job_jump = sys.maxsize
next_gittins_unit = copy.copy(JOBS.gittins_delta)
while (len(JOBS.job_events) + len(JOBS.runnable_jobs))> 0:
if (len(JOBS.job_events) + len(end_events)) == 0:
util.print_fn("This cluster is not large enough to run the job")
break
#decide which is the next event: start or end ?
start_event = None
start_time = sys.maxsize
if len(JOBS.job_events) > 0:
start_event = JOBS.job_events[0]
start_time = start_event['time']
end_event = None
end_time = sys.maxsize
if len(end_events) > 0:
end_event = end_events[0]
end_time = end_event['time']
event_time = sys.maxsize
event = dict()
event['time'] = sys.maxsize
if end_time < start_time:
event_time = end_time
event = end_event
elif end_time > start_time:
event_time = start_time
# event = JOBS.job_events.pop(0)
event = start_event
elif end_time == start_time and end_time != sys.maxsize:
event_time = start_time
# event = JOBS.job_events.pop(0)
event = start_event
event['end_jobs'] = end_events[0]['end_jobs']
assert event_time == event['time']
#decide if job_jump first or (start/end) first