massif/docs/ms-manual.xml

<?xml version="1.0"?> <!-- -*- sgml -*- -->
<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
  "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd">

<chapter id="ms-manual" xreflabel="Massif: a heap profiler">
  <title>Massif: a heap profiler</title>

<para>To use this tool, you must specify
<computeroutput>--tool=massif</computeroutput> on the Valgrind
command line.</para>

<sect1 id="ms-manual.spaceprof" xreflabel="Heap profiling">
<title>Heap profiling</title>

<para>Massif is a heap profiler, i.e. it measures how much heap
memory programs use.  In particular, it can give you information
about:</para>

<itemizedlist>
  <listitem><para>Heap blocks;</para></listitem>
  <listitem><para>Heap administration blocks;</para></listitem>
  <listitem><para>Stack sizes.</para></listitem>
</itemizedlist>

<para>Heap profiling is useful to help you reduce the amount of
memory your program uses.  On modern machines with virtual
memory, this provides the following benefits:</para>

<itemizedlist>
  <listitem><para>It can speed up your program -- a smaller
    program will interact better with your machine's caches,
    avoid paging, and so on.</para></listitem>

  <listitem><para>If your program uses lots of memory, it will
    reduce the chance that it exhausts your machine's swap
    space.</para></listitem>
</itemizedlist>

<para>Also, there are certain space leaks that aren't detected by
traditional leak-checkers, such as Memcheck's.  That's because
the memory isn't ever actually lost -- a pointer remains to it --
but it's not in use.  Programs that have leaks like this can
unnecessarily increase the amount of memory they are using over
time.</para>



<sect2 id="ms-manual.heapprof" 
       xreflabel="Why Use a Heap Profiler?">
<title>Why Use a Heap Profiler?</title>

<para>Everybody knows how useful time profilers are for speeding
up programs.  They are particularly useful because people are
notoriously bad at predicting where are the bottlenecks in their
programs.</para>

<para>But the story is different for heap profilers.  Some
programming languages, particularly lazy functional languages
like <ulink url="http://www.haskell.org">Haskell</ulink>, have
quite sophisticated heap profilers.  But there are few tools as
powerful for profiling C and C++ programs.</para>

<para>Why is this?  Maybe it's because C and C++ programmers must
think that they know where the memory is being allocated.  After
all, you can see all the calls to
<computeroutput>malloc()</computeroutput> and
<computeroutput>new</computeroutput> and
<computeroutput>new[]</computeroutput>, right?  But, in a big
program, do you really know which heap allocations are being
executed, how many times, and how large each allocation is?  Can
you give even a vague estimate of the memory footprint for your
program?  Do you know this for all the libraries your program
uses?  What about administration bytes required by the heap
allocator to track heap blocks -- have you thought about them?
What about the stack?  If you are unsure about any of these
things, maybe you should think about heap profiling.</para>

<para>Massif can tell you these things.</para>

<para>Or maybe it's because it's relatively easy to add basic
heap profiling functionality into a program, to tell you how many
bytes you have allocated for certain objects, or similar.  But
this information might only be simple like total counts for the
whole program's execution.  What about space usage at different
points in the program's execution, for example?  And
reimplementing heap profiling code for each project is a
pain.</para>

<para>Massif can save you this effort.</para>

</sect2>

</sect1>



<sect1 id="ms-manual.using" xreflabel="Using Massif">
<title>Using Massif</title>


<sect2 id="ms-manual.overview" xreflabel="Overview">
<title>Overview</title>

<para>First off, as for normal Valgrind use, you probably want to
compile with debugging info (the
<computeroutput>-g</computeroutput> flag).  But, as opposed to
Memcheck, you probably <command>do</command> want to turn
optimisation on, since you should profile your program as it will
be normally run.</para>

<para>Then, run your program with <computeroutput>valgrind
--tool=massif</computeroutput> in front of the normal command
line invocation.  When the program finishes, Massif will print
summary space statistics.  It also creates a graph representing
the program's heap usage in a file called
<filename>massif.pid.ps</filename>, which can be read by any
PostScript viewer, such as Ghostview.</para>

<para>It also puts detailed information about heap consumption in
a file <filename>massif.pid.txt</filename> (text format) or
<filename>massif.pid.html</filename> (HTML format), where
<emphasis>pid</emphasis> is the program's process id.</para>

</sect2>


<sect2 id="ms-manual.basicresults" xreflabel="Basic Results of Profiling">
<title>Basic Results of Profiling</title>

<para>To gather heap profiling information about the program
<computeroutput>prog</computeroutput>, type:</para>
<screen><![CDATA[
% valgrind --tool=massif prog]]></screen>

<para>The program will execute (slowly).  Upon completion,
summary statistics that look like this will be printed:</para>
<programlisting><![CDATA[
==27519== Total spacetime:   2,258,106 ms.B
==27519== heap:              24.0%
==27519== heap admin:         2.2%
==27519== stack(s):          73.7%]]></programlisting>

<para>All measurements are done in
<emphasis>spacetime</emphasis>, i.e. space (in bytes) multiplied
by time (in milliseconds).  Note that because Massif slows a
program down a lot, the actual spacetime figure is fairly
meaningless; it's the relative values that are
interesting.</para>

<para>Which entries you see in the breakdown depends on the
command line options given.  The above example measures all the
possible parts of memory:</para>

<itemizedlist>
  <listitem><para>Heap: number of words allocated on the heap, via
    <computeroutput>malloc()</computeroutput>,
    <computeroutput>new</computeroutput> and
    <computeroutput>new[]</computeroutput>.</para>
  </listitem>
  <listitem>
    <para>Heap admin: each heap block allocated requires some
    administration data, which lets the allocator track certain
    things about the block.  It is easy to forget about this, and
    if your program allocates lots of small blocks, it can add
    up.  This value is an estimate of the space required for this
    administration data.</para>
  </listitem>
  <listitem>
    <para>Stack(s): the spacetime used by the programs' stack(s).
    (Threaded programs can have multiple stacks.)  This includes
    signal handler stacks.</para>
  </listitem>
</itemizedlist>

</sect2>


<sect2 id="ms-manual.graphs" xreflabel="Spacetime Graphs">
<title>Spacetime Graphs</title>

<para>As well as printing summary information, Massif also
creates a file representing a spacetime graph,
<filename>massif.pid.hp</filename>.  It will produce a file
called <filename>massif.pid.ps</filename>, which can be viewed in
a PostScript viewer.</para>

<para>Massif uses a program called
<computeroutput>hp2ps</computeroutput> to convert the raw data
into the PostScript graph.  It's distributed with Massif, but
came originally from the 
<ulink url="http://haskell.cs.yale.edu/ghc/">Glasgow Haskell
Compiler</ulink>.  You shouldn't need to worry about this at all.
However, if the graph creation fails for any reason, Massif will
tell you, and will leave behind a file named
<filename>massif.pid.hp</filename>, containing the raw heap
profiling data.</para>

<para>Here's an example graph:</para>
<mediaobject id="spacetime-graph">
  <imageobject>
    <imagedata fileref="images/massif-graph-sm.png" format="PNG"/>
  </imageobject>
  <textobject>
    <phrase>Spacetime Graph</phrase>
  </textobject>
</mediaobject>

<para>The graph is broken into several bands.  Most bands
represent a single line of your program that does some heap
allocation; each such band represents all the allocations and
deallocations done from that line.  Up to twenty bands are shown;
less significant allocation sites are merged into "other" and/or
"OTHER" bands.  The accompanying text/HTML file produced by
Massif has more detail about these heap allocation bands.  Then
there are single bands for the stack(s) and heap admin
bytes.</para>

<formalpara>
<title>Note:</title>
<para>it's the height of a band that's important.  Don't let the
ups and downs caused by other bands confuse you.  For example,
the <computeroutput>read_alias_file</computeroutput> band in the
example has the same height all the time it's in existence.</para>
</formalpara>

<para>The triangles on the x-axis show each point at which a
memory census was taken.  These aren't necessarily evenly spread;
Massif only takes a census when memory is allocated or
deallocated.  The time on the x-axis is wallclock time, which is
not ideal because you can get different graphs for different
executions of the same program, due to random OS delays.  But
it's not too bad, and it becomes less of a problem the longer a
program runs.</para>

<para>Massif takes censuses at an appropriate timescale; censuses
take place less frequently as the program runs for longer.  There
is no point having more than 100-200 censuses on a single
graph.</para>

<para>The graphs give a good overview of where your program's
space use comes from, and how that varies over time.  The
accompanying text/HTML file gives a lot more information about
heap use.</para>

</sect2>

</sect1>



<sect1 id="ms-manual.heapdetails" 
       xreflabel="Details of Heap Allocations">
<title>Details of Heap Allocations</title>

<para>The text/HTML file contains information to help interpret
the heap bands of the graph.  It also contains a lot of extra
information about heap allocations that you don't see in the
graph.</para>


<para>Here's part of the information that accompanies the above
graph.</para>

<blockquote>
<literallayout>== 0 ===========================</literallayout>

<para>Heap allocation functions accounted for 50.8% of measured
spacetime</para>

<para>Called from:</para>
<itemizedlist>
  <listitem id="a401767D1"><para>
    <ulink url="#b401767D1">22.1%</ulink>: 0x401767D0:
    _nl_intern_locale_data (in /lib/i686/libc-2.3.2.so)</para>
  </listitem>
  <listitem id="a4017C394"><para>
    <ulink url="#b4017C394">8.6%</ulink>: 0x4017C393:
    read_alias_file (in /lib/i686/libc-2.3.2.so)</para>
  </listitem>
  <listitem>
    <para>... ... <emphasis>(several entries omitted)</emphasis></para>
  </listitem>
  <listitem>
    <para>and 6 other insignificant places</para>
  </listitem>
</itemizedlist>
</blockquote>

<para>The first part shows the total spacetime due to heap
allocations, and the places in the program where most memory was
allocated (Nb: if this program had been compiled with
<computeroutput>-g</computeroutput>, actual line numbers would be
given).  These places are sorted, from most significant to least,
and correspond to the bands seen in the graph.  Insignificant
sites (accounting for less than 0.5% of total spacetime) are
omitted.</para>

<para>That alone can be useful, but often isn't enough.  What if
one of these functions was called from several different places
in the program?  Which one of these is responsible for most of
the memory used?  For
<computeroutput>_nl_intern_locale_data()</computeroutput>, this
question is answered by clicking on the 
<ulink url="#b401767D1">22.1%</ulink> link, which takes us to the
following part of the file:</para>

<blockquote id="b401767D1">
<literallayout>== 1 ===========================</literallayout>

<para>Context accounted for <ulink url="#a401767D1">22.1%</ulink>
of measured spacetime</para>

<para><computeroutput> 0x401767D0: _nl_intern_locale_data (in
/lib/i686/libc-2.3.2.so)</computeroutput></para>

<para>Called from:</para>
<itemizedlist>
  <listitem id="a40176F96"><para>
    <ulink url="#b40176F96">22.1%</ulink>: 0x40176F95:
    _nl_load_locale_from_archive (in
    /lib/i686/libc-2.3.2.so)</para>
  </listitem>
</itemizedlist>
</blockquote>

<para>At this level, we can see all the places from which
<computeroutput>_nl_load_locale_from_archive()</computeroutput>
was called such that it allocated memory at 0x401767D0.  (We can
click on the top <ulink url="#a40176F96">22.1%</ulink> link to go back
to the parent entry.)  At this level, we have moved beyond the
information presented in the graph.  In this case, it is only
called from one place.  We can again follow the link for more
detail, moving to the following part of the file.</para>

<blockquote>
<literallayout>== 2 ===========================</literallayout>
<para id="b40176F96">
Context accounted for <ulink url="#a40176F96">22.1%</ulink> of
measured spacetime</para>

<para><computeroutput> 0x401767D0: _nl_intern_locale_data (in
/lib/i686/libc-2.3.2.so)</computeroutput> <computeroutput>
0x40176F95: _nl_load_locale_from_archive (in
/lib/i686/libc-2.3.2.so)</computeroutput></para>

<para>Called from:</para>
<itemizedlist>
  <listitem id="a40176185">
    <para>22.1%: 0x40176184: _nl_find_locale (in
    /lib/i686/libc-2.3.2.so)</para>
  </listitem>
</itemizedlist>
</blockquote>

<para>In this way we can dig deeper into the call stack, to work
out exactly what sequence of calls led to some memory being
allocated.  At this point, with a call depth of 3, the
information runs out (thus the address of the child entry,
0x40176184, isn't a link).  We could rerun the program with a
greater <computeroutput>--depth</computeroutput> value if we
wanted more information.</para>

<para>Sometimes you will get a code location like this:</para>
<programlisting><![CDATA[
30.8% : 0xFFFFFFFF: ???]]></programlisting>

<para>The code address isn't really 0xFFFFFFFF -- that's
impossible.  This is what Massif does when it can't work out what
the real code address is.</para>

<para>Massif produces this information in a plain text file by
default, or HTML with the
<computeroutput>--format=html</computeroutput> option.  The plain
text version obviously doesn't have the links, but a similar
effect can be achieved by searching on the code addresses.  (In
Vim, the '*' and '#' searches are ideal for this.)</para>


<sect2 id="ms-manual.accuracy" xreflabel="Accuracy">
<title>Accuracy</title>

<para>The information should be pretty accurate.  Some
approximations made might cause some allocation contexts to be
attributed with less memory than they actually allocated, but the
amounts should be miniscule.</para>

<para>The heap admin spacetime figure is an approximation, as
described above.  If anyone knows how to improve its accuracy,
please let us know.</para>

</sect2>

</sect1>


<sect1 id="ms-manual.options" xreflabel="Massif options">
<title>Massif options</title>

<para>Massif-specific options are:</para>

<itemizedlist>

  <listitem>
    <para><computeroutput>--heap=no</computeroutput></para>
    <para><computeroutput>--heap=yes</computeroutput> [default]</para>
    <para>When enabled, profile heap usage in detail.  Without
    it, the <filename>massif.pid.txt</filename> or
    <filename>massif.pid.html</filename> will be very
    short.</para>
  </listitem>

  <listitem>
    <para><computeroutput>--heap-admin=n</computeroutput>
    [default: 8]</para>
    <para>The number of admin bytes per block to use.  This can
    only be an estimate of the average, since it may vary.  The
    allocator used by <computeroutput>glibc</computeroutput>
    requires somewhere between 4--15 bytes per block, depending
    on various factors.  It also requires admin space for freed
    blocks, although Massif does not count this.</para>
  </listitem>

  <listitem>
    <para><computeroutput>--stacks=no</computeroutput></para>
    <para><computeroutput>--stacks=yes</computeroutput> [default]</para>
    <para>When enabled, include stack(s) in the profile.
    Threaded programs can have multiple stacks.</para>
  </listitem>

  <listitem>
    <para><computeroutput>--depth=n</computeroutput>
    [default: 3]</para>
    <para>Depth of call chains to present in the detailed heap
    information.  Increasing it will give more information, but
    Massif will run the program more slowly, using more memory,
    and produce a bigger <computeroutput>.txt</computeroutput> /
    <computeroutput>.hp</computeroutput> file.</para>
  </listitem>

  <listitem>
    <para><computeroutput>--alloc-fn=name</computeroutput></para>
    <para>Specify a function that allocates memory.  This is
    useful for functions that are wrappers to
    <computeroutput>malloc()</computeroutput>, which can fill up
    the context information uselessly (and give very
    uninformative bands on the graph).  Functions specified will
    be ignored in contexts, i.e. treated as though they were
    <computeroutput>malloc()</computeroutput>.  This option can
    be specified multiple times on the command line, to name
    multiple functions.</para>
  </listitem>

  <listitem>
    <para><computeroutput>--format=text</computeroutput> [default]</para>
    <para><computeroutput>--format=html</computeroutput></para>
    <para>Produce the detailed heap information in text or HTML
    format.  The file suffix used will be either
    <computeroutput>.txt</computeroutput> or
    <computeroutput>.html</computeroutput>.</para>
  </listitem>

</itemizedlist>

</sect1>
</chapter>