libstdc++-v3/doc/html/manual/profile_mode_design.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"><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-NS Stylesheets V1.78.1" /><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="profile_mode.html" title="Chapter 19. Profile Mode" /><link rel="prev" href="profile_mode.html" title="Chapter 19. Profile Mode" /><link rel="next" href="profile_mode_api.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="profile_mode_api.html">Next</a></td></tr></table><hr /></div><div class="section"><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>
   3 </p><div class="table"><a id="table.profile_code_loc"></a><p class="title"><strong>Table 19.1. Profile Code Location</strong></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
   4      only included from <code class="code">impl/profiler.h</code>, which is the only
   5      file included from the public headers.</td></tr></tbody></table></div></div><br class="table-break" /><p>
   6 </p><div class="section"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.wrapper"></a>Wrapper Model</h3></div></div></div><p>
   7   In order to get our instrumented library version included instead of the
   8   release one,
   9   we use the same wrapper model as the debug mode.
  10   We subclass entities from the release version.  Wherever
  11   <code class="code">_GLIBCXX_PROFILE</code> is defined, the release namespace is
  12   <code class="code">std::__norm</code>, whereas the profile namespace is
  13   <code class="code">std::__profile</code>.  Using plain <code class="code">std</code> translates
  14   into <code class="code">std::__profile</code>.
  15   </p><p>
  16   Whenever possible, we try to wrap at the public interface level, e.g.,
  17   in <code class="code">unordered_set</code> rather than in <code class="code">hashtable</code>,
  18   in order not to depend on implementation.
  19   </p><p>
  20   Mixing object files built with and without the profile mode must
  21   not affect the program execution.  However, there are no guarantees to
  22   the accuracy of diagnostics when using even a single object not built with
  23   <code class="code">-D_GLIBCXX_PROFILE</code>.
  24   Currently, mixing the profile mode with debug and parallel extensions is
  25   not allowed.  Mixing them at compile time will result in preprocessor errors.
  26   Mixing them at link time is undefined.
  27   </p></div><div class="section"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.instrumentation"></a>Instrumentation</h3></div></div></div><p>
  28   Instead of instrumenting every public entry and exit point,
  29   we chose to add instrumentation on demand, as needed
  30   by individual diagnostics.
  31   The main reason is that some diagnostics require us to extract bits of
  32   internal state that are particular only to that diagnostic.
  33   We plan to formalize this later, after we learn more about the requirements
  34   of several diagnostics.
  35   </p><p>
  36   All the instrumentation points can be switched on and off using
  37   <code class="code">-D[_NO]_GLIBCXX_PROFILE_&lt;diagnostic&gt;</code> options.
  38   With all the instrumentation calls off, there should be negligible
  39   overhead over the release version.  This property is needed to support
  40   diagnostics based on timing of internal operations.  For such diagnostics,
  41   we anticipate turning most of the instrumentation off in order to prevent
  42   profiling overhead from polluting time measurements, and thus diagnostics.
  43   </p><p>
  44   All the instrumentation on/off compile time switches live in
  45   <code class="code">include/profile/profiler.h</code>.
  46   </p></div><div class="section"><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>
  47   For practical reasons, the instrumentation library processes the trace
  48   partially
  49   rather than dumping it to disk in raw form.  Each event is processed when
  50   it occurs.  It is usually attached a cost and it is aggregated into
  51   the database of a specific diagnostic class.  The cost model
  52   is based largely on the standard performance guarantees, but in some
  53   cases we use knowledge about GCC's standard library implementation.
  54   </p><p>
  55   Information is indexed by (1) call stack and (2) instance id or address
  56   to be able to understand and summarize precise creation-use-destruction
  57   dynamic chains.  Although the analysis is sensitive to dynamic instances,
  58   the reports are only sensitive to call context.  Whenever a dynamic instance
  59   is destroyed, we accumulate its effect to the corresponding entry for the
  60   call stack of its constructor location.
  61   </p><p>
  62   For details, see
  63    <a class="link" href="http://dx.doi.org/10.1109/CGO.2009.36" target="_top">paper presented at
  64    CGO 2009</a>.
  65   </p></div><div class="section"><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>
  66   Final analysis takes place offline, and it is based entirely on the
  67   generated trace and debugging info in the application binary.
  68   See section Diagnostics for a list of analysis types that we plan to support.
  69   </p><p>
  70   The input to the analysis is a table indexed by profile type and call stack.
  71   The data type for each entry depends on the profile type.
  72   </p></div><div class="section"><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>
  73   While it is likely that cost models become complex as we get into
  74   more sophisticated analysis, we will try to follow a simple set of rules
  75   at the beginning.
  76   </p><div class="itemizedlist"><ul class="itemizedlist" style="list-style-type: disc; "><li class="listitem"><p><span class="emphasis"><em>Relative benefit estimation:</em></span>
  77   The idea is to estimate or measure the cost of all operations
  78   in the original scenario versus the scenario we advise to switch to.
  79   For instance, when advising to change a vector to a list, an occurrence
  80   of the <code class="code">insert</code> method will generally count as a benefit.
  81   Its magnitude depends on (1) the number of elements that get shifted
  82   and (2) whether it triggers a reallocation.
  83   </p></li><li class="listitem"><p><span class="emphasis"><em>Synthetic measurements:</em></span>
  84   We will measure the relative difference between similar operations on
  85   different containers.  We plan to write a battery of small tests that
  86   compare the times of the executions of similar methods on different
  87   containers.  The idea is to run these tests on the target machine.
  88   If this training phase is very quick, we may decide to perform it at
  89   library initialization time.  The results can be cached on disk and reused
  90   across runs.
  91   </p></li><li class="listitem"><p><span class="emphasis"><em>Timers:</em></span>
  92   We plan to use timers for operations of larger granularity, such as sort.
  93   For instance, we can switch between different sort methods on the fly
  94   and report the one that performs best for each call context.
  95   </p></li><li class="listitem"><p><span class="emphasis"><em>Show stoppers:</em></span>
  96   We may decide that the presence of an operation nullifies the advice.
  97   For instance, when considering switching from <code class="code">set</code> to
  98   <code class="code">unordered_set</code>, if we detect use of operator <code class="code">++</code>,
  99   we will simply not issue the advice, since this could signal that the use
 100   care require a sorted container.</p></li></ul></div></div><div class="section"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.reports"></a>Reports</h3></div></div></div><p>
 101 There are two types of reports.  First, if we recognize a pattern for which
 102 we have a substitute that is likely to give better performance, we print
 103 the advice and estimated performance gain.  The advice is usually associated
 104 to a code position and possibly a call stack.
 105   </p><p>
 106 Second, we report performance characteristics for which we do not have
 107 a clear solution for improvement.  For instance, we can point to the user
 108 the top 10 <code class="code">multimap</code> locations
 109 which have the worst data locality in actual traversals.
 110 Although this does not offer a solution,
 111 it helps the user focus on the key problems and ignore the uninteresting ones.
 112   </p></div><div class="section"><div class="titlepage"><div><div><h3 class="title"><a id="manual.ext.profile_mode.design.testing"></a>Testing</h3></div></div></div><p>
 113   First, we want to make sure we preserve the behavior of the release mode.
 114   You can just type <code class="code">"make check-profile"</code>, which
 115   builds and runs the whole test suite in profile mode.
 116   </p><p>
 117   Second, we want to test the correctness of each diagnostic.
 118   We created a <code class="code">profile</code> directory in the test suite.
 119   Each diagnostic must come with at least two tests, one for false positives
 120   and one for false negatives.
 121   </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="profile_mode_api.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>