OSDN Git Service

2010-04-27 Jonathan Wakely <jwakely.gcc@gmail.com>
[pf3gnuchains/gcc-fork.git] / libstdc++-v3 / doc / html / manual / bk01pt03ch19s02.html
1 <?xml version="1.0" encoding="UTF-8" standalone="no"?>
2 <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
3 <html xmlns="http://www.w3.org/1999/xhtml"><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8" /><title>Design</title><meta name="generator" content="DocBook XSL Stylesheets V1.75.2" /><meta name="keywords" content="&#10;      C++&#10;    , &#10;      library&#10;    , &#10;      profile&#10;    " /><meta name="keywords" content="&#10;      ISO C++&#10;    , &#10;      library&#10;    " /><link rel="home" href="../spine.html" title="The GNU C++ Library Documentation" /><link rel="up" href="profile_mode.html" title="Chapter 19. Profile Mode" /><link rel="prev" href="profile_mode.html" title="Chapter 19. Profile Mode" /><link rel="next" href="bk01pt03ch19s03.html" title="Extensions for Custom Containers" /></head><body><div class="navheader"><table width="100%" summary="Navigation header"><tr><th colspan="3" align="center">Design</th></tr><tr><td width="20%" align="left"><a accesskey="p" href="profile_mode.html">Prev</a> </td><th width="60%" align="center">Chapter 19. Profile Mode</th><td width="20%" align="right"> <a accesskey="n" href="bk01pt03ch19s03.html">Next</a></td></tr></table><hr /></div><div class="sect1" title="Design"><div class="titlepage"><div><div><h2 class="title" style="clear: both"><a id="manual.ext.profile_mode.design"></a>Design</h2></div></div></div><p>
4 </p><div class="table"><a id="id584531"></a><p class="title"><b>Table 19.1. Profile Code Location</b></p><div class="table-contents"><table summary="Profile Code Location" border="1"><colgroup><col align="left" /><col align="left" /></colgroup><thead><tr><th align="left">Code Location</th><th align="left">Use</th></tr></thead><tbody><tr><td align="left"><code class="code">libstdc++-v3/include/std/*</code></td><td align="left">Preprocessor code to redirect to profile extension headers.</td></tr><tr><td align="left"><code class="code">libstdc++-v3/include/profile/*</code></td><td align="left">Profile extension public headers (map, vector, ...).</td></tr><tr><td align="left"><code class="code">libstdc++-v3/include/profile/impl/*</code></td><td align="left">Profile extension internals.  Implementation files are
5      only included from <code class="code">impl/profiler.h</code>, which is the only
6      file included from the public headers.</td></tr></tbody></table></div></div><br class="table-break" /><p>
7 </p><div class="sect2" title="Wrapper Model"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.wrapper"></a>Wrapper Model</h3></div></div></div><p>
8   In order to get our instrumented library version included instead of the
9   release one,
10   we use the same wrapper model as the debug mode.
11   We subclass entities from the release version.  Wherever
12   <code class="code">_GLIBCXX_PROFILE</code> is defined, the release namespace is
13   <code class="code">std::__norm</code>, whereas the profile namespace is
14   <code class="code">std::__profile</code>.  Using plain <code class="code">std</code> translates
15   into <code class="code">std::__profile</code>.
16   </p><p>
17   Whenever possible, we try to wrap at the public interface level, e.g.,
18   in <code class="code">unordered_set</code> rather than in <code class="code">hashtable</code>,
19   in order not to depend on implementation.
20   </p><p>
21   Mixing object files built with and without the profile mode must
22   not affect the program execution.  However, there are no guarantees to
23   the accuracy of diagnostics when using even a single object not built with
24   <code class="code">-D_GLIBCXX_PROFILE</code>.
25   Currently, mixing the profile mode with debug and parallel extensions is
26   not allowed.  Mixing them at compile time will result in preprocessor errors.
27   Mixing them at link time is undefined.
28   </p></div><div class="sect2" title="Instrumentation"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.instrumentation"></a>Instrumentation</h3></div></div></div><p>
29   Instead of instrumenting every public entry and exit point,
30   we chose to add instrumentation on demand, as needed
31   by individual diagnostics.
32   The main reason is that some diagnostics require us to extract bits of
33   internal state that are particular only to that diagnostic.
34   We plan to formalize this later, after we learn more about the requirements
35   of several diagnostics.
36   </p><p>
37   All the instrumentation points can be switched on and off using
38   <code class="code">-D[_NO]_GLIBCXX_PROFILE_&lt;diagnostic&gt;</code> options.
39   With all the instrumentation calls off, there should be negligible
40   overhead over the release version.  This property is needed to support
41   diagnostics based on timing of internal operations.  For such diagnostics,
42   we anticipate turning most of the instrumentation off in order to prevent
43   profiling overhead from polluting time measurements, and thus diagnostics.
44   </p><p>
45   All the instrumentation on/off compile time switches live in
46   <code class="code">include/profile/profiler.h</code>.
47   </p></div><div class="sect2" title="Run Time Behavior"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.rtlib"></a>Run Time Behavior</h3></div></div></div><p>
48   For practical reasons, the instrumentation library processes the trace
49   partially
50   rather than dumping it to disk in raw form.  Each event is processed when
51   it occurs.  It is usually attached a cost and it is aggregated into
52   the database of a specific diagnostic class.  The cost model
53   is based largely on the standard performance guarantees, but in some
54   cases we use knowledge about GCC's standard library implementation.
55   </p><p>
56   Information is indexed by (1) call stack and (2) instance id or address
57   to be able to understand and summarize precise creation-use-destruction
58   dynamic chains.  Although the analysis is sensitive to dynamic instances,
59   the reports are only sensitive to call context.  Whenever a dynamic instance
60   is destroyed, we accumulate its effect to the corresponding entry for the
61   call stack of its constructor location.
62   </p><p>
63   For details, see
64    <a class="ulink" href="http://dx.doi.org/10.1109/CGO.2009.36" target="_top">paper presented at
65    CGO 2009</a>.
66   </p></div><div class="sect2" title="Analysis and Diagnostics"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.analysis"></a>Analysis and Diagnostics</h3></div></div></div><p>
67   Final analysis takes place offline, and it is based entirely on the
68   generated trace and debugging info in the application binary.
69   See section Diagnostics for a list of analysis types that we plan to support.
70   </p><p>
71   The input to the analysis is a table indexed by profile type and call stack.
72   The data type for each entry depends on the profile type.
73   </p></div><div class="sect2" title="Cost Model"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.cost-model"></a>Cost Model</h3></div></div></div><p>
74   While it is likely that cost models become complex as we get into
75   more sophisticated analysis, we will try to follow a simple set of rules
76   at the beginning.
77   </p><div class="itemizedlist"><ul class="itemizedlist" type="disc"><li class="listitem"><p><span class="emphasis"><em>Relative benefit estimation:</em></span>
78   The idea is to estimate or measure the cost of all operations
79   in the original scenario versus the scenario we advise to switch to.
80   For instance, when advising to change a vector to a list, an occurrence
81   of the <code class="code">insert</code> method will generally count as a benefit.
82   Its magnitude depends on (1) the number of elements that get shifted
83   and (2) whether it triggers a reallocation.
84   </p></li><li class="listitem"><p><span class="emphasis"><em>Synthetic measurements:</em></span>
85   We will measure the relative difference between similar operations on
86   different containers.  We plan to write a battery of small tests that
87   compare the times of the executions of similar methods on different
88   containers.  The idea is to run these tests on the target machine.
89   If this training phase is very quick, we may decide to perform it at
90   library initialization time.  The results can be cached on disk and reused
91   across runs.
92   </p></li><li class="listitem"><p><span class="emphasis"><em>Timers:</em></span>
93   We plan to use timers for operations of larger granularity, such as sort.
94   For instance, we can switch between different sort methods on the fly
95   and report the one that performs best for each call context.
96   </p></li><li class="listitem"><p><span class="emphasis"><em>Show stoppers:</em></span>
97   We may decide that the presence of an operation nullifies the advice.
98   For instance, when considering switching from <code class="code">set</code> to
99   <code class="code">unordered_set</code>, if we detect use of operator <code class="code">++</code>,
100   we will simply not issue the advice, since this could signal that the use
101   care require a sorted container.</p></li></ul></div></div><div class="sect2" title="Reports"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.reports"></a>Reports</h3></div></div></div><p>
102 There are two types of reports.  First, if we recognize a pattern for which
103 we have a substitute that is likely to give better performance, we print
104 the advice and estimated performance gain.  The advice is usually associated
105 to a code position and possibly a call stack.
106   </p><p>
107 Second, we report performance characteristics for which we do not have
108 a clear solution for improvement.  For instance, we can point to the user
109 the top 10 <code class="code">multimap</code> locations
110 which have the worst data locality in actual traversals.
111 Although this does not offer a solution,
112 it helps the user focus on the key problems and ignore the uninteresting ones.
113   </p></div><div class="sect2" title="Testing"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.testing"></a>Testing</h3></div></div></div><p>
114   First, we want to make sure we preserve the behavior of the release mode.
115   You can just type <code class="code">"make check-profile"</code>, which
116   builds and runs the whole test suite in profile mode.
117   </p><p>
118   Second, we want to test the correctness of each diagnostic.
119   We created a <code class="code">profile</code> directory in the test suite.
120   Each diagnostic must come with at least two tests, one for false positives
121   and one for false negatives.
122   </p></div></div><div class="navfooter"><hr /><table width="100%" summary="Navigation footer"><tr><td width="40%" align="left"><a accesskey="p" href="profile_mode.html">Prev</a> </td><td width="20%" align="center"><a accesskey="u" href="profile_mode.html">Up</a></td><td width="40%" align="right"> <a accesskey="n" href="bk01pt03ch19s03.html">Next</a></td></tr><tr><td width="40%" align="left" valign="top">Chapter 19. Profile Mode </td><td width="20%" align="center"><a accesskey="h" href="../spine.html">Home</a></td><td width="40%" align="right" valign="top"> Extensions for Custom Containers</td></tr></table></div></body></html>