-
Notifications
You must be signed in to change notification settings - Fork 1.6k
/
perftune.py
executable file
·1889 lines (1521 loc) · 78.1 KB
/
perftune.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
#!/usr/bin/env python3
import abc
import argparse
import distutils.util
import enum
import functools
import glob
import itertools
import logging
import math
import multiprocessing
import os
import pathlib
import pyudev
import re
import shutil
import subprocess
import sys
import urllib.request
import yaml
import platform
import shlex
import psutil
import mmap
import datetime
dry_run_mode = False
def perftune_print(log_msg, *args, **kwargs):
if dry_run_mode:
log_msg = "# " + log_msg
print(log_msg, *args, **kwargs)
def __run_one_command(prog_args, stderr=None, check=True):
proc = subprocess.Popen(prog_args, stdout = subprocess.PIPE, stderr = stderr)
outs, errs = proc.communicate()
outs = str(outs, 'utf-8')
if check and proc.returncode != 0:
raise subprocess.CalledProcessError(returncode=proc.returncode, cmd=" ".join(prog_args), output=outs, stderr=errs)
return outs
def run_one_command(prog_args, stderr=None, check=True):
if dry_run_mode:
print(" ".join([shlex.quote(x) for x in prog_args]))
else:
__run_one_command(prog_args, stderr=stderr, check=check)
def run_read_only_command(prog_args, stderr=None, check=True):
return __run_one_command(prog_args, stderr=stderr, check=check)
def run_hwloc_distrib(prog_args):
"""
Returns a list of strings - each representing a single line of hwloc-distrib output.
"""
return run_read_only_command(['hwloc-distrib'] + prog_args).splitlines()
def run_hwloc_calc(prog_args):
"""
Returns a single string with the result of the execution.
"""
return run_read_only_command(['hwloc-calc'] + prog_args).rstrip()
def run_ethtool(prog_args):
"""
Returns a list of strings - each representing a single line of ethtool output.
"""
return run_read_only_command(['ethtool'] + prog_args).splitlines()
def fwriteln(fname, line, log_message, log_errors=True):
try:
if dry_run_mode:
print("echo {} > {}".format(line, fname))
return
else:
with open(fname, 'w') as f:
f.write(line)
print(log_message)
except:
if log_errors:
print("{}: failed to write into {}: {}".format(log_message, fname, sys.exc_info()))
def readlines(fname):
try:
with open(fname, 'r') as f:
return f.readlines()
except:
print("Failed to read {}: {}".format(fname, sys.exc_info()))
return []
def fwriteln_and_log(fname, line, log_errors=True):
msg = "Writing '{}' to {}".format(line, fname)
fwriteln(fname, line, log_message=msg, log_errors=log_errors)
double_commas_pattern = re.compile(',,')
def set_one_mask(conf_file, mask, log_errors=True):
if not os.path.exists(conf_file):
raise Exception("Configure file to set mask doesn't exist: {}".format(conf_file))
mask = re.sub('0x', '', mask)
while double_commas_pattern.search(mask):
mask = double_commas_pattern.sub(',0,', mask)
msg = "Setting mask {} in {}".format(mask, conf_file)
fwriteln(conf_file, mask, log_message=msg, log_errors=log_errors)
def distribute_irqs(irqs, cpu_mask, log_errors=True):
# If IRQs' list is empty - do nothing
if not irqs:
return
for i, mask in enumerate(run_hwloc_distrib(["{}".format(len(irqs)), '--single', '--restrict', cpu_mask])):
set_one_mask("/proc/irq/{}/smp_affinity".format(irqs[i]), mask, log_errors=log_errors)
def is_process_running(name):
return len(list(filter(lambda ps_line : not re.search('<defunct>', ps_line), run_read_only_command(['ps', '--no-headers', '-C', name], check=False).splitlines()))) > 0
def restart_irqbalance(banned_irqs):
"""
Restart irqbalance if it's running and ban it from moving the IRQs from the
given list.
"""
config_file = '/etc/default/irqbalance'
options_key = 'OPTIONS'
systemd = False
banned_irqs_list = list(banned_irqs)
# If there is nothing to ban - quit
if not banned_irqs_list:
return
# return early if irqbalance is not running
if not is_process_running('irqbalance'):
perftune_print("irqbalance is not running")
return
# If this file exists - this a "new (systemd) style" irqbalance packaging.
# This type of packaging uses IRQBALANCE_ARGS as an option key name, "old (init.d) style"
# packaging uses an OPTION key.
if os.path.exists('/lib/systemd/system/irqbalance.service') or \
os.path.exists('/usr/lib/systemd/system/irqbalance.service'):
options_key = 'IRQBALANCE_ARGS'
systemd = True
if not os.path.exists(config_file):
if os.path.exists('/etc/sysconfig/irqbalance'):
config_file = '/etc/sysconfig/irqbalance'
elif os.path.exists('/etc/conf.d/irqbalance'):
config_file = '/etc/conf.d/irqbalance'
options_key = 'IRQBALANCE_OPTS'
with open('/proc/1/comm', 'r') as comm:
systemd = 'systemd' in comm.read()
else:
perftune_print("Unknown system configuration - not restarting irqbalance!")
perftune_print("You have to prevent it from moving IRQs {} manually!".format(banned_irqs_list))
return
orig_file = "{}.scylla.orig".format(config_file)
# Save the original file
if not dry_run_mode:
if not os.path.exists(orig_file):
print("Saving the original irqbalance configuration is in {}".format(orig_file))
shutil.copyfile(config_file, orig_file)
else:
print("File {} already exists - not overwriting.".format(orig_file))
# Read the config file lines
cfile_lines = open(config_file, 'r').readlines()
# Build the new config_file contents with the new options configuration
perftune_print("Restarting irqbalance: going to ban the following IRQ numbers: {} ...".format(", ".join(banned_irqs_list)))
# Search for the original options line
opt_lines = list(filter(lambda line : re.search(r"^\s*{}".format(options_key), line), cfile_lines))
if not opt_lines:
new_options = "{}=\"".format(options_key)
elif len(opt_lines) == 1:
# cut the last "
new_options = re.sub(r'"\s*$', "", opt_lines[0].rstrip())
opt_lines = opt_lines[0].strip()
else:
raise Exception("Invalid format in {}: more than one lines with {} key".format(config_file, options_key))
for irq in banned_irqs_list:
# prevent duplicate "ban" entries for the same IRQ
opt = f"--banirq={irq}"
patt_str = rf"{opt}\Z|{opt}\s"
if not re.search(patt_str, new_options):
new_options += f" {opt}"
new_options += "\""
if dry_run_mode:
if opt_lines:
print("sed -i 's/^{}/#{}/g' {}".format(options_key, options_key, config_file))
print("echo {} | tee -a {}".format(new_options, config_file))
else:
with open(config_file, 'w') as cfile:
for line in cfile_lines:
if not re.search(r"^\s*{}".format(options_key), line):
cfile.write(line)
cfile.write(new_options + "\n")
if systemd:
perftune_print("Restarting irqbalance via systemctl...")
run_one_command(['systemctl', 'try-restart', 'irqbalance'])
else:
perftune_print("Restarting irqbalance directly (init.d)...")
run_one_command(['/etc/init.d/irqbalance', 'restart'])
def learn_irqs_from_proc_interrupts(pattern, irq2procline):
return [ irq for irq, proc_line in filter(lambda irq_proc_line_pair : re.search(pattern, irq_proc_line_pair[1]), irq2procline.items()) ]
def learn_all_irqs_one(irq_conf_dir, irq2procline, xen_dev_name):
"""
Returns a list of IRQs of a single device.
irq_conf_dir: a /sys/... directory with the IRQ information for the given device
irq2procline: a map of IRQs to the corresponding lines in the /proc/interrupts
xen_dev_name: a device name pattern as it appears in the /proc/interrupts on Xen systems
"""
msi_irqs_dir_name = os.path.join(irq_conf_dir, 'msi_irqs')
# Device uses MSI IRQs
if os.path.exists(msi_irqs_dir_name):
return os.listdir(msi_irqs_dir_name)
irq_file_name = os.path.join(irq_conf_dir, 'irq')
# Device uses INT#x
if os.path.exists(irq_file_name):
return [ line.lstrip().rstrip() for line in open(irq_file_name, 'r').readlines() ]
# No irq file detected
modalias = open(os.path.join(irq_conf_dir, 'modalias'), 'r').readline()
# virtio case
if re.search("^virtio", modalias):
return list(itertools.chain.from_iterable(
map(lambda dirname : learn_irqs_from_proc_interrupts(dirname, irq2procline),
filter(lambda dirname : re.search('virtio', dirname),
itertools.chain.from_iterable([ dirnames for dirpath, dirnames, filenames in os.walk(os.path.join(irq_conf_dir, 'driver')) ])))))
# xen case
if re.search("^xen:", modalias):
return learn_irqs_from_proc_interrupts(xen_dev_name, irq2procline)
return []
def get_irqs2procline_map():
return { line.split(':')[0].lstrip().rstrip() : line for line in open('/proc/interrupts', 'r').readlines() }
class AutodetectError(Exception):
pass
def auto_detect_irq_mask(cpu_mask, cores_per_irq_core):
"""
The logic of auto-detection of what was once a 'mode' is generic and is all about the amount of CPUs and NUMA
nodes that are present and a restricting 'cpu_mask'.
This function implements this logic:
* up to 4 CPU threads: use 'cpu_mask'
* up to 4 CPU cores (on x86 this would translate to 8 CPU threads): use a single CPU thread out of allowed
* up to 16 CPU cores: use a single CPU core out of allowed
* more than 16 CPU cores: use a single CPU core for each 16 CPU cores and distribute them evenly among all
present NUMA nodes. We will also make sure all NUMA nodes have exactly the same amount of IRQ cores.
An AutodetectError exception is raised if 'cpu_mask' is defined in a way that there is a different number of threads
and/or cores among different NUMA nodes. In such a case a user needs to provide
an IRQ CPUs definition explicitly using 'irq_cpu_mask' parameter.
:param cpu_mask: CPU mask that defines which out of present CPUs can be considered for tuning
:param cores_per_irq_core number of cores to allocate a single IRQ core out of, e.g. 6 means allocate a single IRQ
core out of every 6 CPU cores.
:return: CPU mask to bind IRQs to, a.k.a. irq_cpu_mask
"""
cores_key = 'cores'
PUs_key = 'PUs'
# List of NUMA IDs that own CPUs from the given CPU mask
numa_ids_list = run_hwloc_calc(['-I', 'numa', cpu_mask]).split(",")
# Let's calculate number of HTs and cores on each NUMA node belonging to the given CPU set
cores_PUs_per_numa = {} # { <numa_id> : {'cores': <number of cores>, 'PUs': <number of PUs>}}
for n in numa_ids_list:
num_cores = int(run_hwloc_calc(['--restrict', cpu_mask, '--number-of', 'core', f'numa:{n}']))
num_PUs = int(run_hwloc_calc(['--restrict', cpu_mask, '--number-of', 'PU', f'numa:{n}']))
cores_PUs_per_numa[n] = {cores_key: num_cores, PUs_key: num_PUs}
# Let's check if configuration on each NUMA is the same. If it's not then we can't auto-detect the IRQs CPU set
# and a user needs to provide it explicitly
num_cores0 = cores_PUs_per_numa[numa_ids_list[0]][cores_key]
num_PUs0 = cores_PUs_per_numa[numa_ids_list[0]][PUs_key]
for n in numa_ids_list:
if cores_PUs_per_numa[n][cores_key] != num_cores0 or cores_PUs_per_numa[n][PUs_key] != num_PUs0:
raise AutodetectError(f"NUMA{n} has a different configuration from NUMA0 for a given CPU mask {cpu_mask}: "
f"{cores_PUs_per_numa[n][cores_key]}:{cores_PUs_per_numa[n][PUs_key]} vs "
f"{num_cores0}:{num_PUs0}. Auto-detection of IRQ CPUs in not possible. "
f"Please, provide irq_cpu_mask explicitly.")
# Auto-detection of IRQ CPU set is possible - let's get to it!
#
# Total counts for the whole machine
num_cores = int(run_hwloc_calc(['--restrict', cpu_mask, '--number-of', 'core', 'machine:0']))
num_PUs = int(run_hwloc_calc(['--restrict', cpu_mask, '--number-of', 'PU', 'machine:0']))
if num_PUs <= 4:
return cpu_mask
elif num_cores <= 4:
return run_hwloc_calc(['--restrict', cpu_mask, 'PU:0'])
elif num_cores <= cores_per_irq_core:
return run_hwloc_calc(['--restrict', cpu_mask, 'core:0'])
else:
# Big machine.
# Let's allocate a full core out of every cores_per_irq_core cores.
# Let's distribute IRQ cores among present NUMA nodes
num_irq_cores = len(numa_ids_list) * math.ceil(num_cores0 / cores_per_irq_core)
hwloc_args = []
numa_cores_count = {n: 0 for n in numa_ids_list}
added_cores = 0
while added_cores < num_irq_cores:
for numa in numa_ids_list:
hwloc_args.append(f"node:{numa}.core:{numa_cores_count[numa]}")
added_cores += 1
numa_cores_count[numa] += 1
if added_cores >= num_irq_cores:
break
return run_hwloc_calc(['--restrict', cpu_mask] + hwloc_args)
################################################################################
class PerfTunerBase(metaclass=abc.ABCMeta):
def __init__(self, args):
self.__args = args
self.__args.cpu_mask = run_hwloc_calc(['--restrict', self.__args.cpu_mask, 'all'])
self.__mode = None
self.__compute_cpu_mask = None
if self.args.mode:
self.mode = PerfTunerBase.SupportedModes[self.args.mode]
elif args.irq_cpu_mask:
self.irqs_cpu_mask = args.irq_cpu_mask
else:
self.irqs_cpu_mask = auto_detect_irq_mask(self.cpu_mask, self.cores_per_irq_core)
self.__is_aws_i3_nonmetal_instance = None
self.__metadata_token_value = None
self.__metadata_token_time = None
#### Public methods ##########################
class CPUMaskIsZeroException(Exception):
"""Thrown if CPU mask turns out to be zero"""
pass
class SupportedModes(enum.IntEnum):
"""
Modes are ordered from the one that cuts the biggest number of CPUs
from the compute CPUs' set to the one that takes the smallest ('mq' doesn't
cut any CPU from the compute set).
This fact is used when we calculate the 'common quotient' mode out of a
given set of modes (e.g. default modes of different Tuners) - this would
be the smallest among the given modes.
"""
sq_split = 0
sq = 1
mq = 2
# Note: no_irq_restrictions should always have the greatest value in the enum since it's the least restricting mode.
no_irq_restrictions = 9999
@staticmethod
def names():
return PerfTunerBase.SupportedModes.__members__.keys()
@staticmethod
def combine(modes):
"""
:param modes: a set of modes of the PerfTunerBase.SupportedModes type
:return: the mode that is the "common ground" for a given set of modes.
"""
# Perform an explicit cast in order to verify that the values in the 'modes' are compatible with the
# expected PerfTunerBase.SupportedModes type.
return min([PerfTunerBase.SupportedModes(m) for m in modes])
@staticmethod
def cpu_mask_is_zero(cpu_mask):
"""
The cpu_mask is a comma-separated list of 32-bit hex values with possibly omitted zero components,
e.g. 0xffff,,0xffff
We want to estimate if the whole mask is all-zeros.
:param cpu_mask: hwloc-calc generated CPU mask
:return: True if mask is zero, False otherwise
"""
for cur_cpu_mask in cpu_mask.split(','):
if cur_cpu_mask and int(cur_cpu_mask, 16) != 0:
return False
return True
@staticmethod
def compute_cpu_mask_for_mode(mq_mode, cpu_mask):
mq_mode = PerfTunerBase.SupportedModes(mq_mode)
if mq_mode == PerfTunerBase.SupportedModes.sq:
# all but CPU0
compute_cpu_mask = run_hwloc_calc([cpu_mask, '~PU:0'])
elif mq_mode == PerfTunerBase.SupportedModes.sq_split:
# all but CPU0 and its HT siblings
compute_cpu_mask = run_hwloc_calc([cpu_mask, '~core:0'])
elif mq_mode == PerfTunerBase.SupportedModes.mq:
# all available cores
compute_cpu_mask = cpu_mask
elif mq_mode == PerfTunerBase.SupportedModes.no_irq_restrictions:
# all available cores
compute_cpu_mask = cpu_mask
else:
raise Exception("Unsupported mode: {}".format(mq_mode))
if PerfTunerBase.cpu_mask_is_zero(compute_cpu_mask):
raise PerfTunerBase.CPUMaskIsZeroException("Bad configuration mode ({}) and cpu-mask value ({}): this results in a zero-mask for compute".format(mq_mode.name, cpu_mask))
return compute_cpu_mask
@staticmethod
def irqs_cpu_mask_for_mode(mq_mode, cpu_mask):
mq_mode = PerfTunerBase.SupportedModes(mq_mode)
irqs_cpu_mask = 0
if mq_mode != PerfTunerBase.SupportedModes.mq and mq_mode != PerfTunerBase.SupportedModes.no_irq_restrictions:
irqs_cpu_mask = run_hwloc_calc([cpu_mask, "~{}".format(PerfTunerBase.compute_cpu_mask_for_mode(mq_mode, cpu_mask))])
else: # mq_mode == PerfTunerBase.SupportedModes.mq or mq_mode == PerfTunerBase.SupportedModes.no_irq_restrictions
# distribute equally between all available cores
irqs_cpu_mask = cpu_mask
if PerfTunerBase.cpu_mask_is_zero(irqs_cpu_mask):
raise PerfTunerBase.CPUMaskIsZeroException("Bad configuration mode ({}) and cpu-mask value ({}): this results in a zero-mask for IRQs".format(mq_mode.name, cpu_mask))
return irqs_cpu_mask
@property
def mode(self):
"""
Return the configuration mode
"""
return self.__mode
@mode.setter
def mode(self, new_mode):
"""
Set the new configuration mode and recalculate the corresponding masks.
"""
# Make sure the new_mode is of PerfTunerBase.AllowedModes type
self.__mode = PerfTunerBase.SupportedModes(new_mode)
self.__compute_cpu_mask = PerfTunerBase.compute_cpu_mask_for_mode(self.__mode, self.__args.cpu_mask)
self.__irq_cpu_mask = PerfTunerBase.irqs_cpu_mask_for_mode(self.__mode, self.__args.cpu_mask)
@property
def cpu_mask(self):
"""
Return the CPU mask we operate on (the total CPU set)
"""
return self.__args.cpu_mask
@property
def cores_per_irq_core(self):
"""
Return the number of cores we are going to allocate a single IRQ core out of when auto-detecting
"""
return self.__args.cores_per_irq_core
@staticmethod
def min_cores_per_irq_core():
"""
A minimum value of cores_per_irq_core.
We don't allocate a full IRQ core if total number of CPU cores is less or equal to 4.
"""
return 5
@property
def compute_cpu_mask(self):
"""
Return the CPU mask to use for seastar application binding.
"""
return self.__compute_cpu_mask
@property
def irqs_cpu_mask(self):
"""
Return the mask of CPUs used for IRQs distribution.
"""
return self.__irq_cpu_mask
@irqs_cpu_mask.setter
def irqs_cpu_mask(self, new_irq_cpu_mask):
self.__irq_cpu_mask = new_irq_cpu_mask
# Sanity check
if PerfTunerBase.cpu_mask_is_zero(self.__irq_cpu_mask):
raise PerfTunerBase.CPUMaskIsZeroException("Bad configuration: zero IRQ CPU mask is given")
if run_hwloc_calc([self.__irq_cpu_mask]) == run_hwloc_calc([self.cpu_mask]):
# Special case: if IRQ CPU mask is the same as total CPU mask - set a Compute CPU mask to cpu_mask
self.__compute_cpu_mask = self.cpu_mask
else:
# Otherwise, a Compute CPU mask is a CPU mask without IRQ CPU mask bits
self.__compute_cpu_mask = run_hwloc_calc([self.cpu_mask, f"~{self.__irq_cpu_mask}"])
# Sanity check
if PerfTunerBase.cpu_mask_is_zero(self.__compute_cpu_mask):
raise PerfTunerBase.CPUMaskIsZeroException(
f"Bad configuration: cpu_maks:{self.cpu_mask}, irq_cpu_mask:{self.__irq_cpu_mask}: "
f"results in a zero-mask for compute")
@property
def is_aws_i3_non_metal_instance(self):
"""
:return: True if we are running on the AWS i3.nonmetal instance, e.g. i3.4xlarge
"""
if self.__is_aws_i3_nonmetal_instance is None:
self.__check_host_type()
return self.__is_aws_i3_nonmetal_instance
@property
def args(self):
return self.__args
@property
def irqs(self):
return self._get_irqs()
#### "Protected"/Public (pure virtual) methods ###########
@abc.abstractmethod
def tune(self):
pass
@abc.abstractmethod
def _get_irqs(self):
"""
Return the iteratable value with all IRQs to be configured.
"""
pass
#### Private methods ############################
@property
def __ec2_metadata_base_url(self):
return "http://169.254.169.254/latest/"
@property
def __metadata_token(self):
"""
Refresh IMDSv2 session token if it necessary, and return current token
:return: current session token
"""
token_ttl = 21600
update_token = False
if not self.__metadata_token_value:
update_token = True
else:
time_diff = datetime.datetime.now() - self.__metadata_token_time
time_diff_sec = int(time_diff.total_seconds())
if time_diff_sec >= token_ttl - 120:
update_token = True
if update_token:
self.__metadata_token_time = datetime.datetime.now()
req = urllib.request.Request(self.__ec2_metadata_base_url + "api/token", headers={"X-aws-ec2-metadata-token-ttl-seconds": token_ttl}, method="PUT")
with urllib.request.urlopen(req, timeout=0.1) as res:
self.__metadata_token_value = res.read().decode()
return self.__metadata_token_value
def __get_instance_metadata(self, path):
"""
Get a parameter from EC2 Metadata server
:param path: metadata path to access for
:return: received metadata as a string
"""
req = urllib.request.Request(self.__ec2_metadata_base_url + 'meta-data/' + path, headers={"X-aws-ec2-metadata-token": self.__metadata_token})
with urllib.request.urlopen(req, timeout=0.1) as res:
return res.read().decode()
def __check_host_type(self):
"""
Check if we are running on the AWS i3 nonmetal instance.
If yes, set self.__is_aws_i3_nonmetal_instance to True, and to False otherwise.
"""
try:
aws_instance_type = self.__get_instance_metadata('instance-type')
if re.match(r'^i3\.((?!metal)\w)+$', aws_instance_type):
self.__is_aws_i3_nonmetal_instance = True
else:
self.__is_aws_i3_nonmetal_instance = False
return
except (urllib.error.URLError, ConnectionError, TimeoutError):
# Non-AWS case
pass
except:
logging.warning("Unexpected exception while attempting to access AWS meta server: {}".format(sys.exc_info()[0]))
self.__is_aws_i3_nonmetal_instance = False
#################################################
class NetPerfTuner(PerfTunerBase):
def __init__(self, args):
super().__init__(args)
self.nics=args.nics
self.__nic_is_bond_iface = NetPerfTuner.__get_bond_ifaces()
self.__nic_is_vlan_iface = NetPerfTuner.__get_vlan_ifaces()
self.__slaves = self.__learn_slaves()
# check that self.nics contain a HW device or a supported composite interface
self.__check_nics()
# Fetch IRQs related info
self.__get_irqs_info()
#### Public methods ############################
def tune(self):
"""
Tune the networking server configuration.
"""
for nic in self.nics:
if self.nic_is_hw_iface(nic):
perftune_print("Setting a physical interface {}...".format(nic))
self.__setup_one_hw_iface(nic)
else:
perftune_print(f"Setting a {nic} {'bond' if self.nic_is_bond_iface(nic) else 'VLAN'} interface...")
self.__setup_composite_iface(nic)
# Increase the socket listen() backlog
fwriteln_and_log('/proc/sys/net/core/somaxconn', '4096')
# Increase the maximum number of remembered connection requests, which are still
# did not receive an acknowledgment from connecting client.
fwriteln_and_log('/proc/sys/net/ipv4/tcp_max_syn_backlog', '4096')
self.tune_tcp_mem()
def tune_tcp_mem(self):
page_size = mmap.PAGESIZE
total_mem = psutil.virtual_memory().total
# We only tune for physical memory since tcp_mem is virtualized
def to_pages(bytes):
return math.ceil(bytes / page_size)
max = total_mem * self.args.tcp_mem_fraction
fwriteln_and_log('/proc/sys/net/ipv4/tcp_mem', f"{to_pages(max / 2)} {to_pages(max * 2/3)} {to_pages(max)}")
def nic_is_bond_iface(self, nic):
return self.__nic_is_bond_iface.get(nic, False)
def nic_is_vlan_iface(self, nic):
return self.__nic_is_vlan_iface.get(nic, False)
def nic_is_composite_iface(self, nic):
return self.nic_is_bond_iface(nic) or self.nic_is_vlan_iface(nic)
def nic_exists(self, nic):
return self.__iface_exists(nic)
def nic_is_hw_iface(self, nic):
return self.__dev_is_hw_iface(nic)
def slaves(self, nic):
"""
Returns an iterator for all slaves of the nic.
If agrs.nic is not a composite interface an attempt to use the returned iterator
will immediately raise a StopIteration exception - use nic_is_composite_iface() check to avoid this.
"""
return iter(self.__slaves[nic])
#### Protected methods ##########################
def _get_irqs(self):
"""
Returns the iterator for all IRQs that are going to be configured (according to args.nics parameter).
For instance, for a bonding interface that's going to include IRQs of all its slaves.
"""
return itertools.chain.from_iterable(self.__nic2irqs.values())
#### Private methods ############################
def __get_irqs_info(self):
self.__irqs2procline = get_irqs2procline_map()
self.__nic2irqs = self.__learn_irqs()
@property
def __rfs_table_size(self):
return 32768
def __check_nics(self):
"""
Checks that self.nics are supported interfaces
"""
for nic in self.nics:
if not self.nic_exists(nic):
raise Exception("Device {} does not exist".format(nic))
if not self.nic_is_hw_iface(nic) and not self.nic_is_composite_iface(nic):
raise Exception("Not supported virtual device {}".format(nic))
def __get_irqs_one(self, iface):
"""
Returns the list of IRQ numbers for the given interface.
"""
return self.__nic2irqs[iface]
def __setup_rfs(self, iface):
rps_limits = glob.glob("/sys/class/net/{}/queues/*/rps_flow_cnt".format(iface))
one_q_limit = int(self.__rfs_table_size / len(rps_limits))
# If RFS feature is not present - get out
try:
run_one_command(['sysctl', 'net.core.rps_sock_flow_entries'])
except:
return
# Enable RFS
perftune_print("Setting net.core.rps_sock_flow_entries to {}".format(self.__rfs_table_size))
run_one_command(['sysctl', '-w', 'net.core.rps_sock_flow_entries={}'.format(self.__rfs_table_size)])
# Set each RPS queue limit
for rfs_limit_cnt in rps_limits:
msg = "Setting limit {} in {}".format(one_q_limit, rfs_limit_cnt)
fwriteln(rfs_limit_cnt, "{}".format(one_q_limit), log_message=msg)
# Enable/Disable ntuple filtering HW offload on the NIC. This is going to enable/disable aRFS on NICs supporting
# aRFS since ntuple is pre-requisite for an aRFS feature.
# If no explicit configuration has been requested enable ntuple (and thereby aRFS) only in MQ mode.
#
# aRFS acts similar to (SW) RFS: it places a TCP packet on a HW queue that it supposed to be "close" to an
# application thread that sent a packet on the same TCP stream.
#
# For instance if a given TCP stream was sent from CPU3 then the next Rx packet is going to be placed in an Rx
# HW queue which IRQ affinity is set to CPU3 or otherwise to the one with affinity close enough to CPU3.
#
# Read more here: https://access.redhat.com/documentation/en-us/red_hat_enterprise_linux/6/html/performance_tuning_guide/network-acc-rfs
#
# Obviously it would achieve the best result if there is at least one Rx HW queue with an affinity set to each
# application threads that handle TCP.
#
# And, similarly, if we know in advance that there won't be any such HW queue (sq and sq_split modes) - there is
# no sense enabling aRFS.
op = "Enable"
value = 'on'
if (self.args.enable_arfs is None and self.irqs_cpu_mask != self.cpu_mask) or self.args.enable_arfs is False:
op = "Disable"
value = 'off'
ethtool_msg = "{} ntuple filtering HW offload for {}...".format(op, iface)
if dry_run_mode:
perftune_print(ethtool_msg)
run_one_command(['ethtool','-K', iface, 'ntuple', value], stderr=subprocess.DEVNULL)
else:
try:
print("Trying to {} ntuple filtering HW offload for {}...".format(op.lower(), iface), end='')
run_one_command(['ethtool','-K', iface, 'ntuple', value], stderr=subprocess.DEVNULL)
print("ok")
except:
print("not supported")
def __setup_rps(self, iface, mask):
for one_rps_cpus in self.__get_rps_cpus(iface):
set_one_mask(one_rps_cpus, mask)
self.__setup_rfs(iface)
def __setup_xps(self, iface):
xps_cpus_list = glob.glob("/sys/class/net/{}/queues/*/xps_cpus".format(iface))
masks = run_hwloc_distrib(["{}".format(len(xps_cpus_list))])
for i, mask in enumerate(masks):
set_one_mask(xps_cpus_list[i], mask)
def __iface_exists(self, iface):
if len(iface) == 0:
return False
return os.path.exists("/sys/class/net/{}".format(iface))
def __dev_is_hw_iface(self, iface):
return os.path.exists("/sys/class/net/{}/device".format(iface))
@staticmethod
def __get_bond_ifaces():
if not os.path.exists('/sys/class/net/bonding_masters'):
return {}
bond_dict = {}
for line in open('/sys/class/net/bonding_masters', 'r').readlines():
for nic in line.split():
bond_dict[nic] = True
return bond_dict
@staticmethod
def __get_vlan_ifaces():
# Each VLAN interface is going to have a corresponding entry in /proc/net/vlan/ directory
return {pathlib.PurePath(pathlib.Path(f)).name: True
for f in filter(lambda vlan_name: vlan_name != "/proc/net/vlan/config", glob.glob("/proc/net/vlan/*"))}
def __learn_slaves_one(self, nic):
"""
Learn underlying physical devices a given NIC
:param nic: An interface to search slaves for
"""
slaves_list = set()
top_slaves_list = set()
if self.nic_is_bond_iface(nic):
top_slaves_list = set(itertools.chain.from_iterable(
[line.split() for line in open("/sys/class/net/{}/bonding/slaves".format(nic), 'r').readlines()]))
elif self.nic_is_vlan_iface(nic):
# VLAN interfaces have a symbolic link 'lower_<parent_interface_name>' under
# /sys/class/net/<VLAN interface name>.
#
# For example:
#
# lrwxrwxrwx 1 root root 0 Jul 5 18:38 lower_eno1 -> ../../../pci0000:00/0000:00:1f.6/net/eno1/
#
top_slaves_list = set([pathlib.PurePath(pathlib.Path(f).resolve()).name
for f in glob.glob(f"/sys/class/net/{nic}/lower_*")])
# Slaves can be themselves bond or VLAN devices: let's descend (DFS) all the way down to get physical devices.
# Bond slaves can't be VLAN interfaces but VLAN interface parent device can be a bond interface.
# Bond slaves can also be bonds.
# For simplicity let's not discriminate.
for s in top_slaves_list:
if self.nic_is_composite_iface(s):
slaves_list |= self.__learn_slaves_one(s)
else:
slaves_list.add(s)
return slaves_list
def __learn_slaves(self):
"""
Resolve underlying physical devices for interfaces we are requested to configure
"""
slaves_list_per_nic = {}
for nic in self.nics:
current_slaves = self.__learn_slaves_one(nic)
if current_slaves:
slaves_list_per_nic[nic] = list(current_slaves)
return slaves_list_per_nic
def __intel_irq_to_queue_idx(self, irq):
"""
Return the HW queue index for a given IRQ for Intel NICs in order to sort the IRQs' list by this index.
Intel's fast path IRQs have the following name convention:
<bla-bla>-TxRx-<queue index>
Intel NICs also have the IRQ for Flow Director (which is not a regular fast path IRQ) whose name looks like
this:
<bla-bla>:fdir-TxRx-<index>
We want to put the Flow Director's IRQ at the end of the sorted list of IRQs.
:param irq: IRQ number
:return: HW queue index for Intel NICs and sys.maxsize for all other NICs
"""
intel_fp_irq_re = re.compile(r"-TxRx-(\d+)")
fdir_re = re.compile(r"fdir-TxRx-\d+")
m = intel_fp_irq_re.search(self.__irqs2procline[irq])
m1 = fdir_re.search(self.__irqs2procline[irq])
if m and not m1:
return int(m.group(1))
else:
return sys.maxsize
def __mlx_irq_to_queue_idx(self, irq):
"""
Return the HW queue index for a given IRQ for Mellanox NICs in order to sort the IRQs' list by this index.
Mellanox NICs have the IRQ which name looks like
this:
mlx5_comp23
mlx5_comp<index>
or this:
mlx4-6
mlx4-<index>
:param irq: IRQ number
:return: HW queue index for Mellanox NICs and sys.maxsize for all other NICs
"""
mlx5_fp_irq_re = re.compile(r"mlx5_comp(\d+)")
mlx4_fp_irq_re = re.compile(r"mlx4-(\d+)")
m5 = mlx5_fp_irq_re.search(self.__irqs2procline[irq])
if m5:
return int(m5.group(1))
else:
m4 = mlx4_fp_irq_re.search(self.__irqs2procline[irq])
if m4:
return int(m4.group(1))
return sys.maxsize
def __virtio_irq_to_queue_idx(self, irq):
"""
Return the HW queue index for a given IRQ for VIRTIO in order to sort the IRQs' list by this index.
VIRTIO NICs have the IRQ's name that looks like this:
Queue K of a device virtioY, where Y is some integer is comprised of 2 IRQs
with following names:
* Tx IRQ:
virtioY-output.K
* Rx IRQ:
virtioY-input.K
:param irq: IRQ number
:return: HW queue index for VIRTIO fast path IRQ and sys.maxsize for all other IRQs
"""
virtio_fp_re = re.compile(r"virtio\d+-(input|output)\.(\d+)$")
virtio_fp_irq = virtio_fp_re.search(self.__irqs2procline[irq])
if virtio_fp_irq:
return int(virtio_fp_irq.group(2))
return sys.maxsize
def __get_driver_name(self, iface):
"""
:param iface: Interface to check
:return: driver name from ethtool
"""
driver_name = ''
ethtool_i_lines = run_ethtool(['-i', iface])
driver_re = re.compile("driver:")
driver_lines = list(filter(lambda one_line: driver_re.search(one_line), ethtool_i_lines))
if driver_lines:
if len(driver_lines) > 1:
raise Exception("More than one 'driver:' entries in the 'ethtool -i {}' output. Unable to continue.".format(iface))
driver_name = driver_lines[0].split()[1].strip()
return driver_name
def __learn_irqs_one(self, iface):
"""
This is a slow method that is going to read from the system files. Never
use it outside the initialization code. Use __get_irqs_one() instead.
Filter the fast path queues IRQs from the __get_all_irqs_one() result according to the known
patterns.
Right now we know about the following naming convention of the fast path queues vectors:
- Intel: <bla-bla>-TxRx-<bla-bla>
- Broadcom: <bla-bla>-fp-<bla-bla>
- ena: <bla-bla>-Tx-Rx-<bla-bla>
- Mellanox: for mlx4
mlx4-<queue idx>@<bla-bla>
or for mlx5
mlx5_comp<queue idx>@<bla-bla>
- VIRTIO: virtioN-[input|output].D
So, we will try to filter the etries in /proc/interrupts for IRQs we've got from get_all_irqs_one()
according to the patterns above.
If as a result all IRQs are filtered out (if there are no IRQs with the names from the patterns above) then
this means that the given NIC uses a different IRQs naming pattern. In this case we won't filter any IRQ.
Otherwise, we will use only IRQs which names fit one of the patterns above.
For NICs with a limited number of Rx queues the IRQs that handle Rx are going to be at the beginning of the
list.
"""
# filter 'all_irqs' to only reference valid keys from 'irqs2procline' and avoid an IndexError on the 'irqs' search below
all_irqs = set(learn_all_irqs_one("/sys/class/net/{}/device".format(iface), self.__irqs2procline, iface)).intersection(self.__irqs2procline.keys())
fp_irqs_re = re.compile(r"-TxRx-|-fp-|-Tx-Rx-|mlx4-\d+@|mlx5_comp\d+@|virtio\d+-(input|output)")
irqs = sorted(list(filter(lambda irq : fp_irqs_re.search(self.__irqs2procline[irq]), all_irqs)))
if irqs:
irqs.sort(key=self.__get_irq_to_queue_idx_functor(iface))
return irqs
else:
return list(all_irqs)
def __get_irq_to_queue_idx_functor(self, iface):
"""
Get a functor returning a queue index for a given IRQ.
This functor is needed for NICs that are known to not release IRQs when the number of Rx
channels is reduced or have extra IRQs for non-RSS channels.
Therefore, for these NICs we need a functor that would allow us to pick IRQs that belong to channels that are