1 ============================================== 2 Notes on Analysing Behaviour Using Events and 3 ============================================== 4 :Author: Mel Gorman (PCL information heavily b 5 6 1. Introduction 7 =============== 8 9 Tracepoints (see Documentation/trace/tracepoin 10 creating custom kernel modules to register pro 11 tracing infrastructure. 12 13 Simplistically, tracepoints represent importan 14 taken in conjunction with other tracepoints to 15 what is going on within the system. There are 16 gathering and interpreting these events. Lacki 17 this document describes some of the methods th 18 19 This document assumes that debugfs is mounted 20 the appropriate tracing options have been conf 21 assumed that the PCL tool tools/perf has been 22 23 2. Listing Available Events 24 =========================== 25 26 2.1 Standard Utilities 27 ---------------------- 28 29 All possible events are visible from /sys/kern 30 calling:: 31 32 $ find /sys/kernel/tracing/events -type d 33 34 will give a fair indication of the number of e 35 36 2.2 PCL (Performance Counters for Linux) 37 ---------------------------------------- 38 39 Discovery and enumeration of all counters and 40 are available with the perf tool. Getting a li 41 simple case of:: 42 43 $ perf list 2>&1 | grep Tracepoint 44 ext4:ext4_free_inode [Tr 45 ext4:ext4_request_inode [Tr 46 ext4:ext4_allocate_inode [Tr 47 ext4:ext4_write_begin [Tr 48 ext4:ext4_ordered_write_end [Tr 49 [ .... remaining output snipped .... ] 50 51 52 3. Enabling Events 53 ================== 54 55 3.1 System-Wide Event Enabling 56 ------------------------------ 57 58 See Documentation/trace/events.rst for a prope 59 can be enabled system-wide. A short example of 60 to page allocation would look something like:: 61 62 $ for i in `find /sys/kernel/tracing/events 63 64 3.2 System-Wide Event Enabling with SystemTap 65 --------------------------------------------- 66 67 In SystemTap, tracepoints are accessible using 68 call. The following is an example that reports 69 were allocating the pages. 70 :: 71 72 global page_allocs 73 74 probe kernel.trace("mm_page_alloc") { 75 page_allocs[execname()]++ 76 } 77 78 function print_count() { 79 printf ("%-25s %-s\n", "#Pages Allocat 80 foreach (proc in page_allocs-) 81 printf("%-25d %s\n", page_allo 82 printf ("\n") 83 delete page_allocs 84 } 85 86 probe timer.s(5) { 87 print_count() 88 } 89 90 3.3 System-Wide Event Enabling with PCL 91 --------------------------------------- 92 93 By specifying the -a switch and analysing slee 94 for a duration of time can be examined. 95 :: 96 97 $ perf stat -a \ 98 -e kmem:mm_page_alloc -e kmem:mm_page_ 99 -e kmem:mm_page_free_batched \ 100 sleep 10 101 Performance counter stats for 'sleep 10': 102 103 9630 kmem:mm_page_alloc 104 2143 kmem:mm_page_free 105 7424 kmem:mm_page_free_batched 106 107 10.002577764 seconds time elapsed 108 109 Similarly, one could execute a shell and exit 110 at that point. 111 112 3.4 Local Event Enabling 113 ------------------------ 114 115 Documentation/trace/ftrace.rst describes how t 116 basis using set_ftrace_pid. 117 118 3.5 Local Event Enablement with PCL 119 ----------------------------------- 120 121 Events can be activated and tracked for the du 122 basis using PCL such as follows. 123 :: 124 125 $ perf stat -e kmem:mm_page_alloc -e kmem:mm 126 -e kmem:mm_page_free_batched 127 Time: 0.909 128 129 Performance counter stats for './hackbench 130 131 17803 kmem:mm_page_alloc 132 12398 kmem:mm_page_free 133 4827 kmem:mm_page_free_batched 134 135 0.973913387 seconds time elapsed 136 137 4. Event Filtering 138 ================== 139 140 Documentation/trace/ftrace.rst covers in-depth 141 ftrace. Obviously using grep and awk of trace 142 as any script reading trace_pipe. 143 144 5. Analysing Event Variances with PCL 145 ===================================== 146 147 Any workload can exhibit variances between run 148 to know what the standard deviation is. By and 149 performance analyst to do it by hand. In the e 150 occurrences are useful to the performance anal 151 :: 152 153 $ perf stat --repeat 5 -e kmem:mm_page_alloc 154 -e kmem:mm_page_free_b 155 Time: 0.890 156 Time: 0.895 157 Time: 0.915 158 Time: 1.001 159 Time: 0.899 160 161 Performance counter stats for './hackbench 162 163 16630 kmem:mm_page_alloc ( 164 11486 kmem:mm_page_free ( 165 4730 kmem:mm_page_free_batched ( 166 167 0.982653002 seconds time elapsed ( +- 168 169 In the event that some higher-level event is r 170 aggregation of discrete events, then a script 171 172 Using --repeat, it is also possible to view ho 173 time on a system-wide basis using -a and sleep 174 :: 175 176 $ perf stat -e kmem:mm_page_alloc -e kmem:mm 177 -e kmem:mm_page_free_batched \ 178 -a --repeat 10 \ 179 sleep 1 180 Performance counter stats for 'sleep 1' (10 181 182 1066 kmem:mm_page_alloc ( 183 182 kmem:mm_page_free ( 184 890 kmem:mm_page_free_batched ( 185 186 1.002251757 seconds time elapsed ( +- 187 188 6. Higher-Level Analysis with Helper Scripts 189 ============================================ 190 191 When events are enabled the events that are tr 192 /sys/kernel/tracing/trace_pipe in human-readab 193 options exist as well. By post-processing the 194 be gathered on-line as appropriate. Examples o 195 196 - Reading information from /proc for the PID 197 - Deriving a higher-level event from a serie 198 - Calculating latencies between two events 199 200 Documentation/trace/postprocess/trace-pageallo 201 script that can read trace_pipe from STDIN or 202 on-line, it can be interrupted once to generat 203 and twice to exit. 204 205 Simplistically, the script just reads STDIN an 206 also can do more such as 207 208 - Derive high-level events from many low-lev 209 are freed to the main allocator from the p 210 that as one per-CPU drain even though ther 211 for that event 212 - It can aggregate based on PID or individua 213 - In the event memory is getting externally 214 on whether the fragmentation event was sev 215 - When receiving an event about a PID, it ca 216 that if large numbers of events are coming 217 processes, the parent process responsible 218 can be identified 219 220 7. Lower-Level Analysis with PCL 221 ================================ 222 223 There may also be a requirement to identify wh 224 were generating events within the kernel. To b 225 data must be recorded. At the time of writing, 226 :: 227 228 $ perf record -c 1 \ 229 -e kmem:mm_page_alloc -e kmem:mm_page_ 230 -e kmem:mm_page_free_batched \ 231 ./hackbench 10 232 Time: 0.894 233 [ perf record: Captured and wrote 0.733 MB p 234 235 Note the use of '-c 1' to set the event period 236 period is quite high to minimise overhead but 237 very coarse as a result. 238 239 This record outputted a file called perf.data 240 perf report. 241 :: 242 243 $ perf report 244 # Samples: 30922 245 # 246 # Overhead Command Sh 247 # ........ ......... ..................... 248 # 249 87.27% hackbench [vdso] 250 6.85% hackbench /lib/i686/cmov/libc-2 251 2.62% hackbench /lib/ld-2.9.so 252 1.52% perf [vdso] 253 1.22% hackbench ./hackbench 254 0.48% hackbench [kernel] 255 0.02% perf /lib/i686/cmov/libc-2 256 0.01% perf /usr/bin/perf 257 0.01% perf /lib/ld-2.9.so 258 0.00% hackbench /lib/i686/cmov/libpth 259 # 260 # (For more details, try: perf report --sort 261 # 262 263 According to this, the vast majority of events 264 within the VDSO. With simple binaries, this wi 265 take a slightly different example. In the cour 266 noticed that X was generating an insane amount 267 at it: 268 :: 269 270 $ perf record -c 1 -f \ 271 -e kmem:mm_page_alloc -e kmem: 272 -e kmem:mm_page_free_batched \ 273 -p `pidof X` 274 275 This was interrupted after a few seconds and 276 :: 277 278 $ perf report 279 # Samples: 27666 280 # 281 # Overhead Command 282 # ........ ....... ....................... 283 # 284 51.95% Xorg [vdso] 285 47.95% Xorg /opt/gfx-test/lib/libpi 286 0.09% Xorg /lib/i686/cmov/libc-2.9 287 0.01% Xorg [kernel] 288 # 289 # (For more details, try: perf report --sort 290 # 291 292 So, almost half of the events are occurring in 293 symbol: 294 :: 295 296 $ perf report --sort comm,dso,symbol 297 # Samples: 27666 298 # 299 # Overhead Command 300 # ........ ....... ....................... 301 # 302 51.95% Xorg [vdso] 303 47.93% Xorg /opt/gfx-test/lib/libpi 304 0.09% Xorg /lib/i686/cmov/libc-2.9 305 0.01% Xorg /opt/gfx-test/lib/libpi 306 0.01% Xorg [kernel] 307 0.01% Xorg /opt/gfx-test/lib/libpi 308 0.00% Xorg [kernel] 309 310 To see where within the function pixmanFillsse 311 :: 312 313 $ perf annotate pixmanFillsse2 314 [ ... ] 315 0.00 : 34eeb: 0f 18 08 316 : } 317 : 318 : extern __inline void __attribu 319 : _mm_store_si128 (__m128i *__P, 320 : *__P = __B; 321 12.40 : 34eee: 66 0f 7f 80 40 322 0.00 : 34ef5: ff 323 12.40 : 34ef6: 66 0f 7f 80 50 324 0.00 : 34efd: ff 325 12.39 : 34efe: 66 0f 7f 80 60 326 0.00 : 34f05: ff 327 12.67 : 34f06: 66 0f 7f 80 70 328 0.00 : 34f0d: ff 329 12.58 : 34f0e: 66 0f 7f 40 80 330 12.31 : 34f13: 66 0f 7f 40 90 331 12.40 : 34f18: 66 0f 7f 40 a0 332 12.31 : 34f1d: 66 0f 7f 40 b0 333 334 At a glance, it looks like the time is being s 335 the card. Further investigation would be need 336 are being copied around so much but a starting 337 ancient build of libpixmap out of the library 338 forgotten about from months ago!
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.