| 1 |
742 |
jeremybenn |
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
|
| 2 |
|
|
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.1//EN" "http://www.w3.org/TR/xhtml11/DTD/xhtml11.dtd">
|
| 3 |
|
|
<html xmlns="http://www.w3.org/1999/xhtml"><head><title>Chapter 19. Profile Mode</title><meta name="generator" content="DocBook XSL-NS Stylesheets V1.76.1"/><meta name="keywords" content=" C++ , library , profile "/><meta name="keywords" content=" ISO C++ , library "/><meta name="keywords" content=" ISO C++ , runtime , library "/><link rel="home" href="../index.html" title="The GNU C++ Library"/><link rel="up" href="extensions.html" title="Part III. Extensions"/><link rel="prev" href="bk01pt03ch18s05.html" title="Testing"/><link rel="next" href="bk01pt03ch19s02.html" title="Design"/></head><body><div class="navheader"><table width="100%" summary="Navigation header"><tr><th colspan="3" align="center">Chapter 19. Profile Mode</th></tr><tr><td align="left"><a accesskey="p" href="bk01pt03ch18s05.html">Prev</a> </td><th width="60%" align="center">Part III.
|
| 4 |
|
|
Extensions
|
| 5 |
|
|
|
| 6 |
|
|
</th><td align="right"> <a accesskey="n" href="bk01pt03ch19s02.html">Next</a></td></tr></table><hr/></div><div class="chapter" title="Chapter 19. Profile Mode"><div class="titlepage"><div><div><h2 class="title"><a id="manual.ext.profile_mode"/>Chapter 19. Profile Mode</h2></div></div></div><div class="toc"><p><strong>Table of Contents</strong></p><dl><dt><span class="section"><a href="profile_mode.html#manual.ext.profile_mode.intro">Intro</a></span></dt><dd><dl><dt><span class="section"><a href="profile_mode.html#manual.ext.profile_mode.using">Using the Profile Mode</a></span></dt><dt><span class="section"><a href="profile_mode.html#manual.ext.profile_mode.tuning">Tuning the Profile Mode</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s02.html">Design</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.wrapper">Wrapper Model</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.instrumentation">Instrumentation</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.rtlib">Run Time Behavior</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.analysis">Analysis and Diagnostics</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.cost-model">Cost Model</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.reports">Reports</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s02.html#manual.ext.profile_mode.design.testing">Testing</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s03.html">Extensions for Custom Containers</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s04.html">Empirical Cost Model</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s05.html">Implementation Issues</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s05.html#manual.ext.profile_mode.implementation.stack">Stack Traces</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s05.html#manual.ext.profile_mode.implementation.symbols">Symbolization of Instruction Addresses</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s05.html#manual.ext.profile_mode.implementation.concurrency">Concurrency</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s05.html#manual.ext.profile_mode.implementation.stdlib-in-proflib">Using the Standard Library in the Instrumentation Implementation</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s05.html#manual.ext.profile_mode.implementation.malloc-hooks">Malloc Hooks</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s05.html#manual.ext.profile_mode.implementation.construction-destruction">Construction and Destruction of Global Objects</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s06.html">Developer Information</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s06.html#manual.ext.profile_mode.developer.bigpic">Big Picture</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s06.html#manual.ext.profile_mode.developer.howto">How To Add A Diagnostic</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s07.html">Diagnostics</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.template">Diagnostic Template</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.containers">Containers</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.hashtable_too_small">Hashtable Too Small</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.hashtable_too_large">Hashtable Too Large</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.inefficient_hash">Inefficient Hash</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.vector_too_small">Vector Too Small</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.vector_too_large">Vector Too Large</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.vector_to_hashtable">Vector to Hashtable</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.hashtable_to_vector">Hashtable to Vector</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.vector_to_list">Vector to List</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.list_to_vector">List to Vector</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.list_to_slist">List to Forward List (Slist)</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.assoc_ord_to_unord">Ordered to Unordered Associative Container</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.algorithms">Algorithms</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.algorithms.sort">Sort Algorithm Performance</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.locality">Data Locality</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.locality.sw_prefetch">Need Software Prefetch</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.locality.linked">Linked Structure Locality</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.mthread">Multithreaded Data Access</a></span></dt><dd><dl><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.mthread.ddtest">Data Dependence Violations at Container Level</a></span></dt><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.mthread.false_share">False Sharing</a></span></dt></dl></dd><dt><span class="section"><a href="bk01pt03ch19s07.html#manual.ext.profile_mode.analysis.statistics">Statistics</a></span></dt></dl></dd><dt><span class="bibliography"><a href="profile_mode.html#profile_mode.biblio">Bibliography</a></span></dt></dl></div><div class="section" title="Intro"><div class="titlepage"><div><div><h2 class="title"><a id="manual.ext.profile_mode.intro"/>Intro</h2></div></div></div><p>
|
| 7 |
|
|
<span class="emphasis"><em>Goal: </em></span>Give performance improvement advice based on
|
| 8 |
|
|
recognition of suboptimal usage patterns of the standard library.
|
| 9 |
|
|
</p><p>
|
| 10 |
|
|
<span class="emphasis"><em>Method: </em></span>Wrap the standard library code. Insert
|
| 11 |
|
|
calls to an instrumentation library to record the internal state of
|
| 12 |
|
|
various components at interesting entry/exit points to/from the standard
|
| 13 |
|
|
library. Process trace, recognize suboptimal patterns, give advice.
|
| 14 |
|
|
For details, see
|
| 15 |
|
|
<a class="link" href="http://dx.doi.org/10.1109/CGO.2009.36">paper presented at
|
| 16 |
|
|
CGO 2009</a>.
|
| 17 |
|
|
</p><p>
|
| 18 |
|
|
<span class="emphasis"><em>Strengths: </em></span>
|
| 19 |
|
|
</p><div class="itemizedlist"><ul class="itemizedlist"><li class="listitem"><p>
|
| 20 |
|
|
Unintrusive solution. The application code does not require any
|
| 21 |
|
|
modification.
|
| 22 |
|
|
</p></li><li class="listitem"><p> The advice is call context sensitive, thus capable of
|
| 23 |
|
|
identifying precisely interesting dynamic performance behavior.
|
| 24 |
|
|
</p></li><li class="listitem"><p>
|
| 25 |
|
|
The overhead model is pay-per-view. When you turn off a diagnostic class
|
| 26 |
|
|
at compile time, its overhead disappears.
|
| 27 |
|
|
</p></li></ul></div><p>
|
| 28 |
|
|
</p><p>
|
| 29 |
|
|
<span class="emphasis"><em>Drawbacks: </em></span>
|
| 30 |
|
|
</p><div class="itemizedlist"><ul class="itemizedlist"><li class="listitem"><p>
|
| 31 |
|
|
You must recompile the application code with custom options.
|
| 32 |
|
|
</p></li><li class="listitem"><p>You must run the application on representative input.
|
| 33 |
|
|
The advice is input dependent.
|
| 34 |
|
|
</p></li><li class="listitem"><p>
|
| 35 |
|
|
The execution time will increase, in some cases by factors.
|
| 36 |
|
|
</p></li></ul></div><p>
|
| 37 |
|
|
</p><div class="section" title="Using the Profile Mode"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.using"/>Using the Profile Mode</h3></div></div></div><p>
|
| 38 |
|
|
This is the anticipated common workflow for program <code class="code">foo.cc</code>:
|
| 39 |
|
|
</p><pre class="programlisting">
|
| 40 |
|
|
$ cat foo.cc
|
| 41 |
|
|
#include <vector>
|
| 42 |
|
|
int main() {
|
| 43 |
|
|
vector<int> v;
|
| 44 |
|
|
for (int k = 0; k < 1024; ++k) v.insert(v.begin(), k);
|
| 45 |
|
|
}
|
| 46 |
|
|
|
| 47 |
|
|
$ g++ -D_GLIBCXX_PROFILE foo.cc
|
| 48 |
|
|
$ ./a.out
|
| 49 |
|
|
$ cat libstdcxx-profile.txt
|
| 50 |
|
|
vector-to-list: improvement = 5: call stack = 0x804842c ...
|
| 51 |
|
|
: advice = change std::vector to std::list
|
| 52 |
|
|
vector-size: improvement = 3: call stack = 0x804842c ...
|
| 53 |
|
|
: advice = change initial container size from 0 to 1024
|
| 54 |
|
|
</pre><p>
|
| 55 |
|
|
</p><p>
|
| 56 |
|
|
Anatomy of a warning:
|
| 57 |
|
|
</p><div class="itemizedlist"><ul class="itemizedlist"><li class="listitem"><p>
|
| 58 |
|
|
Warning id. This is a short descriptive string for the class
|
| 59 |
|
|
that this warning belongs to. E.g., "vector-to-list".
|
| 60 |
|
|
</p></li><li class="listitem"><p>
|
| 61 |
|
|
Estimated improvement. This is an approximation of the benefit expected
|
| 62 |
|
|
from implementing the change suggested by the warning. It is given on
|
| 63 |
|
|
a log10 scale. Negative values mean that the alternative would actually
|
| 64 |
|
|
do worse than the current choice.
|
| 65 |
|
|
In the example above, 5 comes from the fact that the overhead of
|
| 66 |
|
|
inserting at the beginning of a vector vs. a list is around 1024 * 1024 / 2,
|
| 67 |
|
|
which is around 10e5. The improvement from setting the initial size to
|
| 68 |
|
|
1024 is in the range of 10e3, since the overhead of dynamic resizing is
|
| 69 |
|
|
linear in this case.
|
| 70 |
|
|
</p></li><li class="listitem"><p>
|
| 71 |
|
|
Call stack. Currently, the addresses are printed without
|
| 72 |
|
|
symbol name or code location attribution.
|
| 73 |
|
|
Users are expected to postprocess the output using, for instance, addr2line.
|
| 74 |
|
|
</p></li><li class="listitem"><p>
|
| 75 |
|
|
The warning message. For some warnings, this is static text, e.g.,
|
| 76 |
|
|
"change vector to list". For other warnings, such as the one above,
|
| 77 |
|
|
the message contains numeric advice, e.g., the suggested initial size
|
| 78 |
|
|
of the vector.
|
| 79 |
|
|
</p></li></ul></div><p>
|
| 80 |
|
|
</p><p>Three files are generated. <code class="code">libstdcxx-profile.txt</code>
|
| 81 |
|
|
contains human readable advice. <code class="code">libstdcxx-profile.raw</code>
|
| 82 |
|
|
contains implementation specific data about each diagnostic.
|
| 83 |
|
|
Their format is not documented. They are sufficient to generate
|
| 84 |
|
|
all the advice given in <code class="code">libstdcxx-profile.txt</code>. The advantage
|
| 85 |
|
|
of keeping this raw format is that traces from multiple executions can
|
| 86 |
|
|
be aggregated simply by concatenating the raw traces. We intend to
|
| 87 |
|
|
offer an external utility program that can issue advice from a trace.
|
| 88 |
|
|
<code class="code">libstdcxx-profile.conf.out</code> lists the actual diagnostic
|
| 89 |
|
|
parameters used. To alter parameters, edit this file and rename it to
|
| 90 |
|
|
<code class="code">libstdcxx-profile.conf</code>.
|
| 91 |
|
|
</p><p>Advice is given regardless whether the transformation is valid.
|
| 92 |
|
|
For instance, we advise changing a map to an unordered_map even if the
|
| 93 |
|
|
application semantics require that data be ordered.
|
| 94 |
|
|
We believe such warnings can help users understand the performance
|
| 95 |
|
|
behavior of their application better, which can lead to changes
|
| 96 |
|
|
at a higher abstraction level.
|
| 97 |
|
|
</p></div><div class="section" title="Tuning the Profile Mode"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.tuning"/>Tuning the Profile Mode</h3></div></div></div><p>Compile time switches and environment variables (see also file
|
| 98 |
|
|
profiler.h). Unless specified otherwise, they can be set at compile time
|
| 99 |
|
|
using -D_<name> or by setting variable <name>
|
| 100 |
|
|
in the environment where the program is run, before starting execution.
|
| 101 |
|
|
</p><div class="itemizedlist"><ul class="itemizedlist"><li class="listitem"><p>
|
| 102 |
|
|
<code class="code">_GLIBCXX_PROFILE_NO_<diagnostic></code>:
|
| 103 |
|
|
disable specific diagnostics.
|
| 104 |
|
|
See section Diagnostics for possible values.
|
| 105 |
|
|
(Environment variables not supported.)
|
| 106 |
|
|
</p></li><li class="listitem"><p>
|
| 107 |
|
|
<code class="code">_GLIBCXX_PROFILE_TRACE_PATH_ROOT</code>: set an alternative root
|
| 108 |
|
|
path for the output files.
|
| 109 |
|
|
</p></li><li class="listitem"><p>_GLIBCXX_PROFILE_MAX_WARN_COUNT: set it to the maximum
|
| 110 |
|
|
number of warnings desired. The default value is 10.</p></li><li class="listitem"><p>
|
| 111 |
|
|
<code class="code">_GLIBCXX_PROFILE_MAX_STACK_DEPTH</code>: if set to 0,
|
| 112 |
|
|
the advice will
|
| 113 |
|
|
be collected and reported for the program as a whole, and not for each
|
| 114 |
|
|
call context.
|
| 115 |
|
|
This could also be used in continuous regression tests, where you
|
| 116 |
|
|
just need to know whether there is a regression or not.
|
| 117 |
|
|
The default value is 32.
|
| 118 |
|
|
</p></li><li class="listitem"><p>
|
| 119 |
|
|
<code class="code">_GLIBCXX_PROFILE_MEM_PER_DIAGNOSTIC</code>:
|
| 120 |
|
|
set a limit on how much memory to use for the accounting tables for each
|
| 121 |
|
|
diagnostic type. When this limit is reached, new events are ignored
|
| 122 |
|
|
until the memory usage decreases under the limit. Generally, this means
|
| 123 |
|
|
that newly created containers will not be instrumented until some
|
| 124 |
|
|
live containers are deleted. The default is 128 MB.
|
| 125 |
|
|
</p></li><li class="listitem"><p>
|
| 126 |
|
|
<code class="code">_GLIBCXX_PROFILE_NO_THREADS</code>:
|
| 127 |
|
|
Make the library not use threads. If thread local storage (TLS) is not
|
| 128 |
|
|
available, you will get a preprocessor error asking you to set
|
| 129 |
|
|
-D_GLIBCXX_PROFILE_NO_THREADS if your program is single-threaded.
|
| 130 |
|
|
Multithreaded execution without TLS is not supported.
|
| 131 |
|
|
(Environment variable not supported.)
|
| 132 |
|
|
</p></li><li class="listitem"><p>
|
| 133 |
|
|
<code class="code">_GLIBCXX_HAVE_EXECINFO_H</code>:
|
| 134 |
|
|
This name should be defined automatically at library configuration time.
|
| 135 |
|
|
If your library was configured without <code class="code">execinfo.h</code>, but
|
| 136 |
|
|
you have it in your include path, you can define it explicitly. Without
|
| 137 |
|
|
it, advice is collected for the program as a whole, and not for each
|
| 138 |
|
|
call context.
|
| 139 |
|
|
(Environment variable not supported.)
|
| 140 |
|
|
</p></li></ul></div><p>
|
| 141 |
|
|
</p></div></div><div class="bibliography" title="Bibliography"><div class="titlepage"><div><div><h2 class="title"><a id="profile_mode.biblio"/>Bibliography</h2></div></div></div><div class="biblioentry"><a id="id514403"/><p><span class="citetitle"><em class="citetitle">
|
| 142 |
|
|
Perflint: A Context Sensitive Performance Advisor for C++ Programs
|
| 143 |
|
|
</em>. </span><span class="author"><span class="firstname">Lixia</span> <span class="surname">Liu</span>. </span><span class="author"><span class="firstname">Silvius</span> <span class="surname">Rus</span>. </span><span class="copyright">Copyright © 2009 . </span><span class="publisher"><span class="publishername">
|
| 144 |
|
|
Proceedings of the 2009 International Symposium on Code Generation
|
| 145 |
|
|
and Optimization
|
| 146 |
|
|
. </span></span></p></div></div></div><div class="navfooter"><hr/><table width="100%" summary="Navigation footer"><tr><td align="left"><a accesskey="p" href="bk01pt03ch18s05.html">Prev</a> </td><td align="center"><a accesskey="u" href="extensions.html">Up</a></td><td align="right"> <a accesskey="n" href="bk01pt03ch19s02.html">Next</a></td></tr><tr><td align="left" valign="top">Testing </td><td align="center"><a accesskey="h" href="../index.html">Home</a></td><td align="right" valign="top"> Design</td></tr></table></div></body></html>
|
