This repository has been archived by the owner on Oct 6, 2023. It is now read-only.
forked from iovisor/bcc
-
Notifications
You must be signed in to change notification settings - Fork 4
/
funclatency_example.txt
400 lines (334 loc) · 21 KB
/
funclatency_example.txt
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
Demonstrations of funclatency, the Linux eBPF/bcc version.
Timing the do_sys_open() kernel function until Ctrl-C:
# ./funclatency do_sys_open
Tracing do_sys_open... Hit Ctrl-C to end.
^C
nsecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 0 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 124 |**************** |
4096 -> 8191 : 291 |**************************************|
8192 -> 16383 : 36 |**** |
16384 -> 32767 : 16 |** |
32768 -> 65535 : 8 |* |
65536 -> 131071 : 0 | |
131072 -> 262143 : 0 | |
262144 -> 524287 : 0 | |
524288 -> 1048575 : 0 | |
1048576 -> 2097151 : 0 | |
2097152 -> 4194303 : 1 | |
avg = 13746 nsecs, total: 6543360 nsecs, count: 476
Detaching...
The output shows a histogram of function latency (call time), measured from when
the function began executing (was called) to when it finished (returned).
This example output shows that most of the time, do_sys_open() took between
2048 and 65536 nanoseconds (2 to 65 microseconds). The peak of this distribution
shows 291 calls of between 4096 and 8191 nanoseconds. There was also one
occurrence, an outlier, in the 2 to 4 millisecond range.
How this works: the function entry and return are traced using the kernel kprobe
and kretprobe tracer. Timestamps are collected, the delta time calculated, which
is the bucketized and stored as an in-kernel histogram for efficiency. The
histogram is visible in the output: it's the "count" column; everything else is
decoration. Only the count column is copied to user-level on output. This is an
efficient way to time kernel functions and examine their latency distribution.
Now trace a user function, pthread_mutex_lock in libpthread, to determine if
there is considerable lock contention:
# ./funclatency pthread:pthread_mutex_lock -p $(pidof contentions)
Tracing 1 function for "pthread:pthread_mutex_lock"... Hit Ctrl-C to end.
nsecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 0 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 508967 |****************************************|
4096 -> 8191 : 70072 |***** |
8192 -> 16383 : 27686 |** |
16384 -> 32767 : 5075 | |
32768 -> 65535 : 2318 | |
65536 -> 131071 : 581 | |
131072 -> 262143 : 38 | |
262144 -> 524287 : 5 | |
524288 -> 1048575 : 1 | |
1048576 -> 2097151 : 9 | |
avg = 4317 nsecs, total: 2654426112 nsecs, count: 614752
Detaching...
It seems that most calls to pthread_mutex_lock completed rather quickly (in
under 4us), but there were some cases of considerable contention, sometimes
over a full millisecond.
Run a quick-and-dirty profiler over all the functions in an executable:
# ./funclatency /home/user/primes:* -p $(pidof primes) -F
Tracing 15 functions for "/home/user/primes:*"... Hit Ctrl-C to end.
^C
Function = is_prime [6556]
nsecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 0 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 1495322 |****************************************|
4096 -> 8191 : 95744 |** |
8192 -> 16383 : 9926 | |
16384 -> 32767 : 3070 | |
32768 -> 65535 : 1415 | |
65536 -> 131071 : 112 | |
131072 -> 262143 : 9 | |
262144 -> 524287 : 3 | |
524288 -> 1048575 : 0 | |
1048576 -> 2097151 : 8 | |
Function = insert_result [6556]
nsecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 0 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 111047 |****************************************|
4096 -> 8191 : 3998 |* |
8192 -> 16383 : 720 | |
16384 -> 32767 : 238 | |
32768 -> 65535 : 106 | |
65536 -> 131071 : 5 | |
131072 -> 262143 : 4 | |
avg = 3404 nsecs, total: 5862276096 nsecs, count: 1721727
Detaching...
From the results, we can see that the is_prime function has something resembling
an exponential distribution -- very few primes take a very long time to test,
while most numbers are verified as prime or composite in less than 4us. The
insert_result function exhibits a similar phenomenon, likely due to contention
over a shared results container.
Now vfs_read() is traced, and a microseconds histogram printed:
# ./funclatency -u vfs_read
Tracing vfs_read... Hit Ctrl-C to end.
^C
usecs : count distribution
0 -> 1 : 1143 |**************************************|
2 -> 3 : 420 |************* |
4 -> 7 : 159 |***** |
8 -> 15 : 295 |********* |
16 -> 31 : 25 | |
32 -> 63 : 5 | |
64 -> 127 : 1 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 1 | |
2048 -> 4095 : 0 | |
4096 -> 8191 : 5 | |
8192 -> 16383 : 0 | |
16384 -> 32767 : 0 | |
32768 -> 65535 : 0 | |
65536 -> 131071 : 7 | |
131072 -> 262143 : 7 | |
262144 -> 524287 : 3 | |
524288 -> 1048575 : 7 | |
avg = 4229 nsecs, total: 8789145 nsecs, count: 2078
Detaching...
This shows a bimodal distribution. Many vfs_read() calls were faster than 15
microseconds, however, there was also a small handful between 65 milliseconds
and 1 second, seen at the bottom of the table. These are likely network reads
from SSH, waiting on interactive keystrokes.
Tracing do_nanosleep() in milliseconds:
# ./funclatency -m do_nanosleep
Tracing do_nanosleep... Hit Ctrl-C to end.
^C
msecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 0 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 328 |**************************************|
1024 -> 2047 : 0 | |
2048 -> 4095 : 0 | |
4096 -> 8191 : 32 |*** |
8192 -> 16383 : 0 | |
16384 -> 32767 : 0 | |
32768 -> 65535 : 2 | |
avg = 1510 nsecs, total: 546816 nsecs, count: 326
Detaching...
This looks like it has found threads that are sleeping every 1, 5, and 60
seconds.
An interval can be provided using -i, and timestamps added using -T. For
example, tracing vfs_read() latency in milliseconds and printing output
every 5 seconds:
# ./funclatency -mTi 5 vfs_read
Tracing vfs_read... Hit Ctrl-C to end.
20:10:08
msecs : count distribution
0 -> 1 : 1500 |*************************************+|
2 -> 3 : 3 | |
4 -> 7 : 1 | |
8 -> 15 : 2 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 4 | |
128 -> 255 : 3 | |
256 -> 511 : 1 | |
512 -> 1023 : 7 | |
avg = 5 nsecs, total: 8259 nsecs, count: 1521
20:10:13
msecs : count distribution
0 -> 1 : 1251 |*************************************+|
2 -> 3 : 3 | |
4 -> 7 : 2 | |
8 -> 15 : 0 | |
16 -> 31 : 2 | |
32 -> 63 : 3 | |
64 -> 127 : 5 | |
128 -> 255 : 5 | |
256 -> 511 : 3 | |
512 -> 1023 : 6 | |
1024 -> 2047 : 2 | |
avg = 9 nsecs, total: 11736 nsecs, count: 1282
20:10:18
msecs : count distribution
0 -> 1 : 1265 |*************************************+|
2 -> 3 : 0 | |
4 -> 7 : 5 | |
8 -> 15 : 9 | |
16 -> 31 : 7 | |
32 -> 63 : 1 | |
64 -> 127 : 2 | |
128 -> 255 : 3 | |
256 -> 511 : 5 | |
512 -> 1023 : 5 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 1 | |
avg = 8 nsecs, total: 11219 nsecs, count: 1303
^C
20:10:20
msecs : count distribution
0 -> 1 : 249 |*************************************+|
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 1 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 1 | |
avg = 4 nsecs, total: 1029 nsecs, count: 251
Detaching...
A single process can be traced, which filters in-kernel for efficiency. Here,
the vfs_read() function is timed as milliseconds for PID 17064, which is a
bash shell:
# ./funclatency -mp 17064 vfs_read
Tracing vfs_read... Hit Ctrl-C to end.
^C
msecs : count distribution
0 -> 1 : 1 |** |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 1 |** |
16 -> 31 : 2 |***** |
32 -> 63 : 0 | |
64 -> 127 : 13 |**************************************|
128 -> 255 : 10 |***************************** |
256 -> 511 : 4 |*********** |
avg = 153 nsecs, total: 4765 nsecs, count: 31
Detaching...
The distribution between 64 and 511 milliseconds shows keystroke latency.
The -F option can be used to print a histogram per function. Eg:
# ./funclatency -uF 'vfs_r*'
Tracing 5 functions for "vfs_r*"... Hit Ctrl-C to end.
^C
Function = vfs_read
usecs : count distribution
0 -> 1 : 1044 |****************************************|
2 -> 3 : 383 |************** |
4 -> 7 : 76 |** |
8 -> 15 : 41 |* |
16 -> 31 : 26 | |
32 -> 63 : 0 | |
64 -> 127 : 1 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 4 | |
4096 -> 8191 : 2 | |
8192 -> 16383 : 0 | |
16384 -> 32767 : 0 | |
32768 -> 65535 : 2 | |
65536 -> 131071 : 5 | |
131072 -> 262143 : 5 | |
262144 -> 524287 : 3 | |
524288 -> 1048575 : 7 | |
Function = vfs_rename
usecs : count distribution
0 -> 1 : 2 |**** |
2 -> 3 : 2 |**** |
4 -> 7 : 2 |**** |
8 -> 15 : 0 | |
16 -> 31 : 6 |************* |
32 -> 63 : 18 |****************************************|
avg = 5087 nsecs, total: 8287001 nsecs, count: 1629
Detaching...
USAGE message:
# ./funclatency -h
usage: funclatency [-h] [-p PID] [-i INTERVAL] [-T] [-u] [-m] [-F] [-r] [-v]
pattern
Time functions and print latency as a histogram
positional arguments:
pattern search expression for functions
optional arguments:
-h, --help show this help message and exit
-p PID, --pid PID trace this PID only
-i INTERVAL, --interval INTERVAL
summary interval, in seconds
-d DURATION, --duration DURATION
total duration of trace, in seconds
-T, --timestamp include timestamp on output
-u, --microseconds microsecond histogram
-m, --milliseconds millisecond histogram
-F, --function show a separate histogram per function
-r, --regexp use regular expressions. Default is "*" wildcards
only.
-l LEVEL, --level LEVEL
set the level of nested or recursive functions
-v, --verbose print the BPF program (for debugging purposes)
examples:
./funclatency do_sys_open # time the do_sys_open() kernel function
./funclatency c:read # time the read() C library function
./funclatency -u vfs_read # time vfs_read(), in microseconds
./funclatency -m do_nanosleep # time do_nanosleep(), in milliseconds
./funclatency -i 2 -d 10 c:open # output every 2 seconds, for duration 10s
./funclatency -mTi 5 vfs_read # output every 5 seconds, with timestamps
./funclatency -p 181 vfs_read # time process 181 only
./funclatency 'vfs_fstat*' # time both vfs_fstat() and vfs_fstatat()
./funclatency 'c:*printf' # time the *printf family of functions
./funclatency -F 'vfs_r*' # show one histogram per matched function