• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1<!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML//EN">
2<HTML>
3
4<HEAD>
5  <link rel="stylesheet" href="designstyle.css">
6  <title>Gperftools CPU Profiler</title>
7</HEAD>
8
9<BODY>
10
11<p align=right>
12  <i>Last modified
13  <script type=text/javascript>
14    var lm = new Date(document.lastModified);
15    document.write(lm.toDateString());
16  </script></i>
17</p>
18
19<p>This is the CPU profiler we use at Google.  There are three parts
20to using it: linking the library into an application, running the
21code, and analyzing the output.</p>
22
23<p>On the off-chance that you should need to understand it, the CPU
24profiler data file format is documented separately,
25<a href="cpuprofile-fileformat.html">here</a>.
26
27
28<H1>Linking in the Library</H1>
29
30<p>To install the CPU profiler into your executable, add
31<code>-lprofiler</code> to the link-time step for your executable.
32(It's also probably possible to add in the profiler at run-time using
33<code>LD_PRELOAD</code>, e.g.
34<code>% env LD_PRELOAD="/usr/lib/libprofiler.so" &lt;binary&gt;</code>,
35but this isn't necessarily recommended.)</p>
36
37<p>This does <i>not</i> turn on CPU profiling; it just inserts the
38code.  For that reason, it's practical to just always link
39<code>-lprofiler</code> into a binary while developing; that's what we
40do at Google.  (However, since any user can turn on the profiler by
41setting an environment variable, it's not necessarily recommended to
42install profiler-linked binaries into a production, running
43system.)</p>
44
45
46<H1>Running the Code</H1>
47
48<p>There are several alternatives to actually turn on CPU profiling
49for a given run of an executable:</p>
50
51<ol>
52  <li> <p>Define the environment variable CPUPROFILE to the filename
53       to dump the profile to.  For instance, if you had a version of
54       <code>/bin/ls</code> that had been linked against libprofiler,
55       you could run:</p>
56       <pre>% env CPUPROFILE=ls.prof /bin/ls</pre>
57
58  <li> <p>In your code, bracket the code you want profiled in calls to
59       <code>ProfilerStart()</code> and <code>ProfilerStop()</code>.
60       (These functions are declared in <code>&lt;gperftools/profiler.h&gt;</code>.)
61       <code>ProfilerStart()</code> will take
62       the profile-filename as an argument.</p>
63</ol>
64
65<p>In Linux 2.6 and above, profiling works correctly with threads,
66automatically profiling all threads.  In Linux 2.4, profiling only
67profiles the main thread (due to a kernel bug involving itimers and
68threads).  Profiling works correctly with sub-processes: each child
69process gets its own profile with its own name (generated by combining
70CPUPROFILE with the child's process id).</p>
71
72<p>For security reasons, CPU profiling will not write to a file -- and
73is thus not usable -- for setuid programs.</p>
74
75<p>See the include-file <code>gperftools/profiler.h</code> for
76advanced-use functions, including <code>ProfilerFlush()</code> and
77<code>ProfilerStartWithOptions()</code>.</p>
78
79
80<H2>Modifying Runtime Behavior</H2>
81
82<p>You can more finely control the behavior of the CPU profiler via
83environment variables.</p>
84
85<table frame=box rules=sides cellpadding=5 width=100%>
86
87<tr valign=top>
88  <td><code>CPUPROFILE_FREQUENCY=<i>x</i></code></td>
89  <td>default: 100</td>
90  <td>
91    How many interrupts/second the cpu-profiler samples.
92  </td>
93</tr>
94
95<tr valign=top>
96  <td><code>CPUPROFILE_REALTIME=1</code></td>
97  <td>default: [not set]</td>
98  <td>
99    If set to any value (including 0 or the empty string), use
100    ITIMER_REAL instead of ITIMER_PROF to gather profiles.  In
101    general, ITIMER_REAL is not as accurate as ITIMER_PROF, and also
102    interacts badly with use of alarm(), so prefer ITIMER_PROF unless
103    you have a reason prefer ITIMER_REAL.
104  </td>
105</tr>
106
107</table>
108
109
110<h1><a name="pprof">Analyzing the Output</a></h1>
111
112<p><code>pprof</code> is the script used to analyze a profile.  It has
113many output modes, both textual and graphical.  Some give just raw
114numbers, much like the <code>-pg</code> output of <code>gcc</code>,
115and others show the data in the form of a dependency graph.</p>
116
117<p>pprof <b>requires</b> <code>perl5</code> to be installed to run.
118It also requires <code>dot</code> to be installed for any of the
119graphical output routines, and <code>gv</code> to be installed for
120<code>--gv</code> mode (described below).
121</p>
122
123<p>Here are some ways to call pprof.  These are described in more
124detail below.</p>
125
126<pre>
127% pprof /bin/ls ls.prof
128                       Enters "interactive" mode
129% pprof --text /bin/ls ls.prof
130                       Outputs one line per procedure
131% pprof --gv /bin/ls ls.prof
132                       Displays annotated call-graph via 'gv'
133% pprof --gv --focus=Mutex /bin/ls ls.prof
134                       Restricts to code paths including a .*Mutex.* entry
135% pprof --gv --focus=Mutex --ignore=string /bin/ls ls.prof
136                       Code paths including Mutex but not string
137% pprof --list=getdir /bin/ls ls.prof
138                       (Per-line) annotated source listing for getdir()
139% pprof --disasm=getdir /bin/ls ls.prof
140                       (Per-PC) annotated disassembly for getdir()
141% pprof --text localhost:1234
142                       Outputs one line per procedure for localhost:1234
143% pprof --callgrind /bin/ls ls.prof
144                       Outputs the call information in callgrind format
145</pre>
146
147
148<h3>Analyzing Text Output</h3>
149
150<p>Text mode has lines of output that look like this:</p>
151<pre>
152       14   2.1%  17.2%       58   8.7% std::_Rb_tree::find
153</pre>
154
155<p>Here is how to interpret the columns:</p>
156<ol>
157  <li> Number of profiling samples in this function
158  <li> Percentage of profiling samples in this function
159  <li> Percentage of profiling samples in the functions printed so far
160  <li> Number of profiling samples in this function and its callees
161  <li> Percentage of profiling samples in this function and its callees
162  <li> Function name
163</ol>
164
165<h3>Analyzing Callgrind Output</h3>
166
167<p>Use <a href="http://kcachegrind.sourceforge.net">kcachegrind</a> to
168analyze your callgrind output:</p>
169<pre>
170% pprof --callgrind /bin/ls ls.prof > ls.callgrind
171% kcachegrind ls.callgrind
172</pre>
173
174<p>The cost is specified in 'hits', i.e. how many times a function
175appears in the recorded call stack information. The 'calls' from
176function a to b record how many times function b was found in the
177stack traces directly below function a.</p>
178
179<p>Tip: if you use a debug build the output will include file and line
180number information and kcachegrind will show an annotated source
181code view.</p>
182
183<h3>Node Information</h3>
184
185<p>In the various graphical modes of pprof, the output is a call graph
186annotated with timing information, like so:</p>
187
188<A HREF="pprof-test-big.gif">
189<center><table><tr><td>
190   <img src="pprof-test.gif">
191</td></tr></table></center>
192</A>
193
194<p>Each node represents a procedure.  The directed edges indicate
195caller to callee relations.  Each node is formatted as follows:</p>
196
197<center><pre>
198Class Name
199Method Name
200local (percentage)
201<b>of</b> cumulative (percentage)
202</pre></center>
203
204<p>The last one or two lines contains the timing information.  (The
205profiling is done via a sampling method, where by default we take 100
206samples a second.  Therefor one unit of time in the output corresponds
207to about 10 milliseconds of execution time.) The "local" time is the
208time spent executing the instructions directly contained in the
209procedure (and in any other procedures that were inlined into the
210procedure).  The "cumulative" time is the sum of the "local" time and
211the time spent in any callees.  If the cumulative time is the same as
212the local time, it is not printed.</p>
213
214<p>For instance, the timing information for test_main_thread()
215indicates that 155 units (about 1.55 seconds) were spent executing the
216code in <code>test_main_thread()</code> and 200 units were spent while
217executing <code>test_main_thread()</code> and its callees such as
218<code>snprintf()</code>.</p>
219
220<p>The size of the node is proportional to the local count.  The
221percentage displayed in the node corresponds to the count divided by
222the total run time of the program (that is, the cumulative count for
223<code>main()</code>).</p>
224
225<h3>Edge Information</h3>
226
227<p>An edge from one node to another indicates a caller to callee
228relationship.  Each edge is labelled with the time spent by the callee
229on behalf of the caller.  E.g, the edge from
230<code>test_main_thread()</code> to <code>snprintf()</code> indicates
231that of the 200 samples in <code>test_main_thread()</code>, 37 are
232because of calls to <code>snprintf()</code>.</p>
233
234<p>Note that <code>test_main_thread()</code> has an edge to
235<code>vsnprintf()</code>, even though <code>test_main_thread()</code>
236doesn't call that function directly.  This is because the code was
237compiled with <code>-O2</code>; the profile reflects the optimized
238control flow.</p>
239
240<h3>Meta Information</h3>
241
242<p>The top of the display should contain some meta information
243like:</p>
244<pre>
245      /tmp/profiler2_unittest
246      Total samples: 202
247      Focusing on: 202
248      Dropped nodes with &lt;= 1 abs(samples)
249      Dropped edges with &lt;= 0 samples
250</pre>
251
252<p>This section contains the name of the program, and the total
253samples collected during the profiling run.  If the
254<code>--focus</code> option is on (see the <a href="#focus">Focus</a>
255section below), the legend also contains the number of samples being
256shown in the focused display.  Furthermore, some unimportant nodes and
257edges are dropped to reduce clutter.  The characteristics of the
258dropped nodes and edges are also displayed in the legend.</p>
259
260<h3><a name=focus>Focus and Ignore</a></h3>
261
262<p>You can ask pprof to generate a display focused on a particular
263piece of the program.  You specify a regular expression.  Any portion
264of the call-graph that is on a path which contains at least one node
265matching the regular expression is preserved.  The rest of the
266call-graph is dropped on the floor.  For example, you can focus on the
267<code>vsnprintf()</code> libc call in <code>profiler2_unittest</code>
268as follows:</p>
269
270<pre>
271% pprof --gv --focus=vsnprintf /tmp/profiler2_unittest test.prof
272</pre>
273<A HREF="pprof-vsnprintf-big.gif">
274<center><table><tr><td>
275   <img src="pprof-vsnprintf.gif">
276</td></tr></table></center>
277</A>
278
279<p>Similarly, you can supply the <code>--ignore</code> option to
280ignore samples that match a specified regular expression.  E.g., if
281you are interested in everything except calls to
282<code>snprintf()</code>, you can say:</p>
283<pre>
284% pprof --gv --ignore=snprintf /tmp/profiler2_unittest test.prof
285</pre>
286
287
288<h3>Interactive mode</a></h3>
289
290<p>By default -- if you don't specify any flags to the contrary --
291pprof runs in interactive mode.  At the <code>(pprof)</code> prompt,
292you can run many of the commands described above.  You can type
293<code>help</code> for a list of what commands are available in
294interactive mode.</p>
295
296<h3><a name=options>pprof Options</a></h3>
297
298For a complete list of pprof options, you can run <code>pprof
299--help</code>.
300
301<h4>Output Type</h4>
302
303<p>
304<center>
305<table frame=box rules=sides cellpadding=5 width=100%>
306<tr valign=top>
307  <td><code>--text</code></td>
308  <td>
309    Produces a textual listing.  (Note: If you have an X display, and
310    <code>dot</code> and <code>gv</code> installed, you will probably
311    be happier with the <code>--gv</code> output.)
312  </td>
313</tr>
314<tr valign=top>
315  <td><code>--gv</code></td>
316  <td>
317    Generates annotated call-graph, converts to postscript, and
318    displays via gv (requres <code>dot</code> and <code>gv</code> be
319    installed).
320  </td>
321</tr>
322<tr valign=top>
323  <td><code>--dot</code></td>
324  <td>
325    Generates the annotated call-graph in dot format and
326    emits to stdout (requres <code>dot</code> be installed).
327  </td>
328</tr>
329<tr valign=top>
330  <td><code>--ps</code></td>
331  <td>
332    Generates the annotated call-graph in Postscript format and
333    emits to stdout (requres <code>dot</code> be installed).
334  </td>
335</tr>
336<tr valign=top>
337  <td><code>--pdf</code></td>
338  <td>
339    Generates the annotated call-graph in PDF format and emits to
340    stdout (requires <code>dot</code> and <code>ps2pdf</code> be
341    installed).
342  </td>
343</tr>
344<tr valign=top>
345  <td><code>--gif</code></td>
346  <td>
347    Generates the annotated call-graph in GIF format and
348    emits to stdout (requres <code>dot</code> be installed).
349  </td>
350</tr>
351<tr valign=top>
352  <td><code>--list=&lt;<i>regexp</i>&gt;</code></td>
353  <td>
354    <p>Outputs source-code listing of routines whose
355    name matches &lt;regexp&gt;.  Each line
356    in the listing is annotated with flat and cumulative
357    sample counts.</p>
358
359    <p>In the presence of inlined calls, the samples
360    associated with inlined code tend to get assigned
361    to a line that follows the location of the
362    inlined call.  A more precise accounting can be
363    obtained by disassembling the routine using the
364    --disasm flag.</p>
365  </td>
366</tr>
367<tr valign=top>
368  <td><code>--disasm=&lt;<i>regexp</i>&gt;</code></td>
369  <td>
370    Generates disassembly of routines that match
371    &lt;regexp&gt;, annotated with flat and
372    cumulative sample counts and emits to stdout.
373  </td>
374</tr>
375</table>
376</center>
377
378<h4>Reporting Granularity</h4>
379
380<p>By default, pprof produces one entry per procedure.  However you can
381use one of the following options to change the granularity of the
382output.  The <code>--files</code> option seems to be particularly
383useless, and may be removed eventually.</p>
384
385<center>
386<table frame=box rules=sides cellpadding=5 width=100%>
387<tr valign=top>
388  <td><code>--addresses</code></td>
389  <td>
390     Produce one node per program address.
391  </td>
392</tr>
393  <td><code>--lines</code></td>
394  <td>
395     Produce one node per source line.
396  </td>
397</tr>
398  <td><code>--functions</code></td>
399  <td>
400     Produce one node per function (this is the default).
401  </td>
402</tr>
403  <td><code>--files</code></td>
404  <td>
405     Produce one node per source file.
406  </td>
407</tr>
408</table>
409</center>
410
411<h4>Controlling the Call Graph Display</h4>
412
413<p>Some nodes and edges are dropped to reduce clutter in the output
414display.  The following options control this effect:</p>
415
416<center>
417<table frame=box rules=sides cellpadding=5 width=100%>
418<tr valign=top>
419  <td><code>--nodecount=&lt;n&gt;</code></td>
420  <td>
421    This option controls the number of displayed nodes.  The nodes
422    are first sorted by decreasing cumulative count, and then only
423    the top N nodes are kept.  The default value is 80.
424  </td>
425</tr>
426<tr valign=top>
427  <td><code>--nodefraction=&lt;f&gt;</code></td>
428  <td>
429    This option provides another mechanism for discarding nodes
430    from the display.  If the cumulative count for a node is
431    less than this option's value multiplied by the total count
432    for the profile, the node is dropped.  The default value
433    is 0.005; i.e. nodes that account for less than
434    half a percent of the total time are dropped.  A node
435    is dropped if either this condition is satisfied, or the
436    --nodecount condition is satisfied.
437  </td>
438</tr>
439<tr valign=top>
440  <td><code>--edgefraction=&lt;f&gt;</code></td>
441  <td>
442    This option controls the number of displayed edges.  First of all,
443    an edge is dropped if either its source or destination node is
444    dropped.  Otherwise, the edge is dropped if the sample
445    count along the edge is less than this option's value multiplied
446    by the total count for the profile.  The default value is
447    0.001; i.e., edges that account for less than
448    0.1% of the total time are dropped.
449  </td>
450</tr>
451<tr valign=top>
452  <td><code>--focus=&lt;re&gt;</code></td>
453  <td>
454    This option controls what region of the graph is displayed
455    based on the regular expression supplied with the option.
456    For any path in the callgraph, we check all nodes in the path
457    against the supplied regular expression.  If none of the nodes
458    match, the path is dropped from the output.
459  </td>
460</tr>
461<tr valign=top>
462  <td><code>--ignore=&lt;re&gt;</code></td>
463  <td>
464    This option controls what region of the graph is displayed
465    based on the regular expression supplied with the option.
466    For any path in the callgraph, we check all nodes in the path
467    against the supplied regular expression.  If any of the nodes
468    match, the path is dropped from the output.
469  </td>
470</tr>
471</table>
472</center>
473
474<p>The dropped edges and nodes account for some count mismatches in
475the display.  For example, the cumulative count for
476<code>snprintf()</code> in the first diagram above was 41.  However
477the local count (1) and the count along the outgoing edges (12+1+20+6)
478add up to only 40.</p>
479
480
481<h1>Caveats</h1>
482
483<ul>
484  <li> If the program exits because of a signal, the generated profile
485       will be <font color=red>incomplete, and may perhaps be
486       completely empty</font>.
487  <li> The displayed graph may have disconnected regions because
488       of the edge-dropping heuristics described above.
489  <li> If the program linked in a library that was not compiled
490       with enough symbolic information, all samples associated
491       with the library may be charged to the last symbol found
492       in the program before the library.  This will artificially
493       inflate the count for that symbol.
494  <li> If you run the program on one machine, and profile it on
495       another, and the shared libraries are different on the two
496       machines, the profiling output may be confusing: samples that
497       fall within  shared libaries may be assigned to arbitrary
498       procedures.
499  <li> If your program forks, the children will also be profiled
500       (since they inherit the same CPUPROFILE setting).  Each process
501       is profiled separately; to distinguish the child profiles from
502       the parent profile and from each other, all children will have
503       their process-id appended to the CPUPROFILE name.
504  <li> Due to a hack we make to work around a possible gcc bug, your
505       profiles may end up named strangely if the first character of
506       your CPUPROFILE variable has ascii value greater than 127.
507       This should be exceedingly rare, but if you need to use such a
508       name, just set prepend <code>./</code> to your filename:
509       <code>CPUPROFILE=./&Auml;gypten</code>.
510</ul>
511
512
513<hr>
514<address>Sanjay Ghemawat<br>
515<!-- Created: Tue Dec 19 10:43:14 PST 2000 -->
516<!-- hhmts start -->
517Last modified: Fri May  9 14:41:29 PDT 2008
518<!-- hhmts end -->
519</address>
520</BODY>
521</HTML>
522