libstdc++-v3/doc/html/manual/bk01pt03ch19s02.html

   1 <html><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8"><title>Design</title><meta name="generator" content="DocBook XSL-NS Stylesheets V1.76.1"><meta name="keywords" content="
   2       C++
   3     ,
   4       library
   5     ,
   6       profile
   7     "><meta name="keywords" content="
   8       ISO C++
   9     ,
  10       library
  11     "><meta name="keywords" content="
  12       ISO C++
  13     ,
  14       runtime
  15     ,
  16       library
  17     "><link rel="home" href="../index.html" title="The GNU C++ Library"><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 bgcolor="white" text="black" link="#0000FF" vlink="#840084" alink="#0000FF"><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="section" title="Design"><div class="titlepage"><div><div><h2 class="title" style="clear: both"><a name="manual.ext.profile_mode.design"></a>Design</h2></div></div></div><p>
  18 </p><div class="table"><a name="id635111"></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" class="c1"><col align="left" class="c2"></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
  19      only included from <code class="code">impl/profiler.h</code>, which is the only
  20      file included from the public headers.</td></tr></tbody></table></div></div><br class="table-break"><p>
  21 </p><div class="section" title="Wrapper Model"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.wrapper"></a>Wrapper Model</h3></div></div></div><p>
  22   In order to get our instrumented library version included instead of the
  23   release one,
  24   we use the same wrapper model as the debug mode.
  25   We subclass entities from the release version.  Wherever
  26   <code class="code">_GLIBCXX_PROFILE</code> is defined, the release namespace is
  27   <code class="code">std::__norm</code>, whereas the profile namespace is
  28   <code class="code">std::__profile</code>.  Using plain <code class="code">std</code> translates
  29   into <code class="code">std::__profile</code>.
  30   </p><p>
  31   Whenever possible, we try to wrap at the public interface level, e.g.,
  32   in <code class="code">unordered_set</code> rather than in <code class="code">hashtable</code>,
  33   in order not to depend on implementation.
  34   </p><p>
  35   Mixing object files built with and without the profile mode must
  36   not affect the program execution.  However, there are no guarantees to
  37   the accuracy of diagnostics when using even a single object not built with
  38   <code class="code">-D_GLIBCXX_PROFILE</code>.
  39   Currently, mixing the profile mode with debug and parallel extensions is
  40   not allowed.  Mixing them at compile time will result in preprocessor errors.
  41   Mixing them at link time is undefined.
  42   </p></div><div class="section" title="Instrumentation"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.instrumentation"></a>Instrumentation</h3></div></div></div><p>
  43   Instead of instrumenting every public entry and exit point,
  44   we chose to add instrumentation on demand, as needed
  45   by individual diagnostics.
  46   The main reason is that some diagnostics require us to extract bits of
  47   internal state that are particular only to that diagnostic.
  48   We plan to formalize this later, after we learn more about the requirements
  49   of several diagnostics.
  50   </p><p>
  51   All the instrumentation points can be switched on and off using
  52   <code class="code">-D[_NO]_GLIBCXX_PROFILE_&lt;diagnostic&gt;</code> options.
  53   With all the instrumentation calls off, there should be negligible
  54   overhead over the release version.  This property is needed to support
  55   diagnostics based on timing of internal operations.  For such diagnostics,
  56   we anticipate turning most of the instrumentation off in order to prevent
  57   profiling overhead from polluting time measurements, and thus diagnostics.
  58   </p><p>
  59   All the instrumentation on/off compile time switches live in
  60   <code class="code">include/profile/profiler.h</code>.
  61   </p></div><div class="section" title="Run Time Behavior"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.rtlib"></a>Run Time Behavior</h3></div></div></div><p>
  62   For practical reasons, the instrumentation library processes the trace
  63   partially
  64   rather than dumping it to disk in raw form.  Each event is processed when
  65   it occurs.  It is usually attached a cost and it is aggregated into
  66   the database of a specific diagnostic class.  The cost model
  67   is based largely on the standard performance guarantees, but in some
  68   cases we use knowledge about GCC's standard library implementation.
  69   </p><p>
  70   Information is indexed by (1) call stack and (2) instance id or address
  71   to be able to understand and summarize precise creation-use-destruction
  72   dynamic chains.  Although the analysis is sensitive to dynamic instances,
  73   the reports are only sensitive to call context.  Whenever a dynamic instance
  74   is destroyed, we accumulate its effect to the corresponding entry for the
  75   call stack of its constructor location.
  76   </p><p>
  77   For details, see
  78    <a class="link" href="http://dx.doi.org/10.1109/CGO.2009.36" target="_top">paper presented at
  79    CGO 2009</a>.
  80   </p></div><div class="section" title="Analysis and Diagnostics"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.analysis"></a>Analysis and Diagnostics</h3></div></div></div><p>
  81   Final analysis takes place offline, and it is based entirely on the
  82   generated trace and debugging info in the application binary.
  83   See section Diagnostics for a list of analysis types that we plan to support.
  84   </p><p>
  85   The input to the analysis is a table indexed by profile type and call stack.
  86   The data type for each entry depends on the profile type.
  87   </p></div><div class="section" title="Cost Model"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.cost-model"></a>Cost Model</h3></div></div></div><p>
  88   While it is likely that cost models become complex as we get into
  89   more sophisticated analysis, we will try to follow a simple set of rules
  90   at the beginning.
  91   </p><div class="itemizedlist"><ul class="itemizedlist" type="disc"><li class="listitem"><p><span class="emphasis"><em>Relative benefit estimation:</em></span>
  92   The idea is to estimate or measure the cost of all operations
  93   in the original scenario versus the scenario we advise to switch to.
  94   For instance, when advising to change a vector to a list, an occurrence
  95   of the <code class="code">insert</code> method will generally count as a benefit.
  96   Its magnitude depends on (1) the number of elements that get shifted
  97   and (2) whether it triggers a reallocation.
  98   </p></li><li class="listitem"><p><span class="emphasis"><em>Synthetic measurements:</em></span>
  99   We will measure the relative difference between similar operations on
 100   different containers.  We plan to write a battery of small tests that
 101   compare the times of the executions of similar methods on different
 102   containers.  The idea is to run these tests on the target machine.
 103   If this training phase is very quick, we may decide to perform it at
 104   library initialization time.  The results can be cached on disk and reused
 105   across runs.
 106   </p></li><li class="listitem"><p><span class="emphasis"><em>Timers:</em></span>
 107   We plan to use timers for operations of larger granularity, such as sort.
 108   For instance, we can switch between different sort methods on the fly
 109   and report the one that performs best for each call context.
 110   </p></li><li class="listitem"><p><span class="emphasis"><em>Show stoppers:</em></span>
 111   We may decide that the presence of an operation nullifies the advice.
 112   For instance, when considering switching from <code class="code">set</code> to
 113   <code class="code">unordered_set</code>, if we detect use of operator <code class="code">++</code>,
 114   we will simply not issue the advice, since this could signal that the use
 115   care require a sorted container.</p></li></ul></div></div><div class="section" title="Reports"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.reports"></a>Reports</h3></div></div></div><p>
 116 There are two types of reports.  First, if we recognize a pattern for which
 117 we have a substitute that is likely to give better performance, we print
 118 the advice and estimated performance gain.  The advice is usually associated
 119 to a code position and possibly a call stack.
 120   </p><p>
 121 Second, we report performance characteristics for which we do not have
 122 a clear solution for improvement.  For instance, we can point to the user
 123 the top 10 <code class="code">multimap</code> locations
 124 which have the worst data locality in actual traversals.
 125 Although this does not offer a solution,
 126 it helps the user focus on the key problems and ignore the uninteresting ones.
 127   </p></div><div class="section" title="Testing"><div class="titlepage"><div><div><h3 class="title"><a name="manual.ext.profile_mode.design.testing"></a>Testing</h3></div></div></div><p>
 128   First, we want to make sure we preserve the behavior of the release mode.
 129   You can just type <code class="code">"make check-profile"</code>, which
 130   builds and runs the whole test suite in profile mode.
 131   </p><p>
 132   Second, we want to test the correctness of each diagnostic.
 133   We created a <code class="code">profile</code> directory in the test suite.
 134   Each diagnostic must come with at least two tests, one for false positives
 135   and one for false negatives.
 136   </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="../index.html">Home</a></td><td width="40%" align="right" valign="top"> Extensions for Custom Containers</td></tr></table></div></body></html>