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 Stylesheets Vsnapshot" /><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 class="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_<diagnostic></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="https://ieeexplore.ieee.org/document/4907670/" 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> 101There are two types of reports. First, if we recognize a pattern for which 102we have a substitute that is likely to give better performance, we print 103the advice and estimated performance gain. The advice is usually associated 104to a code position and possibly a call stack. 105 </p><p> 106Second, we report performance characteristics for which we do not have 107a clear solution for improvement. For instance, we can point to the user 108the top 10 <code class="code">multimap</code> locations 109which have the worst data locality in actual traversals. 110Although this does not offer a solution, 111it 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>