valgrind/dhat/docs/dh-manual.xml

<?xml version="1.0"?> <!-- -*- sgml -*- -->
<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.5//EN"
          "http://www.oasis-open.org/docbook/xml/4.5/docbookx.dtd"
[ <!ENTITY % vg-entities SYSTEM "../../docs/xml/vg-entities.xml"> %vg-entities; ]>


<chapter id="dh-manual" 
         xreflabel="DHAT: a dynamic heap analysis tool">
  <title>DHAT: a dynamic heap analysis tool</title>

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



<sect1 id="dh-manual.overview" xreflabel="Overview">
<title>Overview</title>

<para>DHAT is primarily a tool for examining how programs use their heap
allocations.</para>

<para>It tracks the allocated blocks, and inspects every memory access
to find which block, if any, it is to. It presents, on a program point
basis, information about these blocks such as sizes, lifetimes, numbers of
reads and writes, and read and write patterns.</para>

<para>Using this information it is possible to identify program points with
the following characteristics:</para>

<itemizedlist>

  <listitem><para>potential process-lifetime leaks: blocks allocated
   by the point just accumulate, and are freed only at the end of the
   run.</para></listitem>

 <listitem><para>excessive turnover: points which chew through a lot
  of heap, even if it is not held onto for very long</para></listitem>

 <listitem><para>excessively transient: points which allocate very
 short lived blocks</para></listitem>

 <listitem><para>useless or underused allocations: blocks which are
  allocated but not completely filled in, or are filled in but not
  subsequently read.</para></listitem>

 <listitem><para>blocks with inefficient layout -- areas never
  accessed, or with hot fields scattered throughout the
  block.</para></listitem>
</itemizedlist>

<para>As with the Massif heap profiler, DHAT measures program progress
by counting instructions, and so presents all age/time related figures
as instruction counts. This sounds a little odd at first, but it
makes runs repeatable in a way which is not possible if CPU time is
used.</para>

<para>DHAT also has support for copy profiling and ad hoc profiling. These are
described below.</para>

</sect1>



<sect1 id="dh-manual.profile" xreflabel="Using DHAT">
<title>Using DHAT</title>

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

<para>Second, you need to run your program under DHAT to gather the profiling
information. You might need to reduce the <option>--num-callers</option> value
to get reasonably-sized output files, especially if you are profiling a large
program; some trial and error might be needed to find a good value.</para>

<para>Finally, you need to use DHAT's viewer (in a web browser) to get a
detailed presentation of that information.</para>


<sect2 id="dh-manual.running-DHAT" xreflabel="Running DHAT">
<title>Running DHAT</title>

<para>To run DHAT on a program <filename>prog</filename>, run:</para>
<screen><![CDATA[
valgrind --tool=dhat prog
]]></screen>

<para>The program will execute (slowly). Upon completion, summary statistics
that look like this will be printed:</para>

<programlisting><![CDATA[
==11514== Total:     823,849,731 bytes in 3,929,133 blocks
==11514== At t-gmax: 133,485,082 bytes in 436,521 blocks
==11514== At t-end:  258,002 bytes in 2,129 blocks
==11514== Reads:     2,807,182,810 bytes
==11514== Writes:    1,149,617,086 bytes
]]></programlisting>

<para>The first line shows how many heap blocks and bytes were allocated over
the entire execution.</para>

<para>The second line shows how many heap blocks and bytes were alive at
<computeroutput>t-gmax</computeroutput>, i.e. the time when the heap size
reached its global maximum (as measured in bytes).</para>

<para>The third line shows how many heap blocks and bytes were alive at
<computeroutput>t-end</computeroutput>, i.e. the end of execution. In other
words, how many blocks and bytes were not explicitly freed. </para>

<para>The fourth and fifth lines show how many bytes within heap blocks were
read and written during the entire execution. </para>

<para>These lines are moderately interesting at best. More useful information
can be seen with DHAT's viewer.</para>

</sect2>


<sect2 id="dh-manual.outputfile" xreflabel="Output File">
<title>Output File</title>

<para>As well as printing summary information, DHAT also writes more detailed
profiling information to a file. By default this file is named
<filename>dhat.out.&lt;pid&gt;</filename> (where
<filename>&lt;pid&gt;</filename> is the program's process ID), but its name can
be changed with the <option>--dhat-out-file</option> option. This file is JSON,
and intended to be viewed by DHAT's viewer, which is described in the next
section.</para>

<para>The default <computeroutput>.&lt;pid&gt;</computeroutput> suffix on the
output file name serves two purposes. Firstly, it means you don't have to
rename old log files that you don't want to overwrite. Secondly, and more
importantly, it allows correct profiling with the
<option>--trace-children=yes</option> option of programs that spawn child
processes.</para>

<para>The output file can be big, many megabytes for large applications
built with full debugging information.</para>

</sect2>

</sect1>



<sect1 id="dh-manual.viewer" xreflabel="DHAT's viewer">
<title>DHAT's Viewer</title>

<para>DHAT's viewer can be run in a web browser by loading the file
<computeroutput>dh_view.html</computeroutput>. Use the "Load" button to choose
a DHAT output file to view.</para>

<para>If loading takes a long time, it might be worth re-running DHAT with a
smaller <option>--num-callers</option> value to reduce the stack depths,
because this can significantly reduce the size of DHAT's output files.</para>


<sect2 id="dh-output-header"><title>The Output Header</title>

<para>The first part of the output shows the mode, program command and process
ID. For example:</para>

<programlisting><![CDATA[
Invocation {
  Mode:    heap
  Command: /home/njn/moz/rust0/build/x86_64-unknown-linux-gnu/stage2/bin/rustc --crate-name tuple_stress src/main.rs
  PID:     18816
}
]]></programlisting>

<para>The second part of the output shows the
<computeroutput>t-gmax</computeroutput> and
<computeroutput>t-end</computeroutput> values again. For example:</para>

<programlisting><![CDATA[
Times {
  t-gmax: 8,138,210,673 instrs (86.92% of program duration)
  t-end:  9,362,544,994 instrs
}
]]></programlisting>

</sect2>


<sect2 id="dh-ap-tree"><title>The PP Tree</title>

<para>The third part of the output is the largest and most interesting part,
showing the program point (PP) tree.</para>


<sect3 id="dh-structure"><title>Structure</title>

<para>The following image shows a screenshot of part of a PP
tree. The font is very small because this screenshot is intended to
demonstrate the high-level structure of the tree rather than the
details within the text. (It is also slightly out-of-date, and doesn't quite
match the current output produced by DHAT's viewer.)</para>

<graphic fileref="images/dh-tree.png" scalefit="1"/>

<para>Like any tree, it has a root node, leaf nodes, and non-leaf nodes. The
structure of the tree is shown by the lines connecting nodes. Child nodes are
beneath their parent and indented one level.</para>

<para>The sub-trees beneath a non-leaf node can be collapsed or expanded by
clicking on the node. It is useful to collapse sub-trees that you aren't
interested in.</para>

<para>Colours are meaningful, and are intended to ease tree navigation, but the
information they represent is also present within the text. (This means that
colour-blind users are not denied any information.)</para>

<para>Each leaf node is coloured green. Each non-leaf node is coloured blue
and has a down arrow (<computeroutput>▼</computeroutput>) next to it when
its sub-tree is expanded. Each non-leaf node is coloured yellow and has a
left arrow (<computeroutput>▶</computeroutput>) next to it when its sub-tree
is collapsed.</para>

<para>The shade of green, blue or yellow used for a node indicate its
significance. Darker shades represent greater significance (in terms of bytes
or blocks).</para>

<para>Note that the entire output is text, even the arrows and lines connecting
nodes. This means you can copy and paste any part of the output easily into an
email, bug report, etc.</para>

</sect3>


<sect3 id="dh-root-node"><title>The Root Node</title>

<para>The root node looks like this:</para>

<programlisting><![CDATA[
PP 1/1 (25 children) {
  Total:     1,355,253,987 bytes (100%, 67,454.81/Minstr) in 5,943,417 blocks (100%, 295.82/Minstr), avg size 228.03 bytes, avg lifetime 3,134,692,250.67 instrs (15.6% of program duration)
  At t-gmax: 423,930,307 bytes (100%) in 1,575,682 blocks (100%), avg size 269.05 bytes
  At t-end:  258,002 bytes (100%) in 2,129 blocks (100%), avg size 121.18 bytes
  Reads:     5,478,606,988 bytes (100%, 272,685.7/Minstr), 4.04/byte
  Writes:    2,040,294,800 bytes (100%, 101,551.22/Minstr), 1.51/byte
  Allocated at {
    #0: [root]
  }
}
]]></programlisting>

<para>The root node covers the entire execution. The information is a superset
of the information shown when DHAT ran, adding details such as allocation
rates, average block sizes, block lifetimes, and read and write ratios. The
next example will explain these in more detail.</para>

</sect3>


<sect3 id="dh-interior-nodes"><title>Interior Nodes</title>

<para>PP nodes further down the tree show information about a subset of
allocations. For example:</para>

<programlisting><![CDATA[
PP 1.1/25 (2 children) {
  Total:     54,533,440 bytes (4.02%, 2,714.28/Minstr) in 458,839 blocks (7.72%, 22.84/Minstr), avg size 118.85 bytes, avg lifetime 1,127,259,403.64 instrs (5.61% of program duration)
  At t-gmax: 0 bytes (0%) in 0 blocks (0%), avg size 0 bytes
  At t-end:  0 bytes (0%) in 0 blocks (0%), avg size 0 bytes
  Reads:     15,993,012 bytes (0.29%, 796.02/Minstr), 0.29/byte
  Writes:    20,974,752 bytes (1.03%, 1,043.97/Minstr), 0.38/byte
  Allocated at {
    #1: 0x95CACC9: alloc (alloc.rs:72)
    #2: 0x95CACC9: alloc (alloc.rs:148)
    #3: 0x95CACC9: reserve_internal<syntax::tokenstream::TokenStream,alloc::alloc::Global> (raw_vec.rs:669)
    #4: 0x95CACC9: reserve<syntax::tokenstream::TokenStream,alloc::alloc::Global> (raw_vec.rs:492)
    #5: 0x95CACC9: reserve<syntax::tokenstream::TokenStream> (vec.rs:460)
    #6: 0x95CACC9: push<syntax::tokenstream::TokenStream> (vec.rs:989)
    #7: 0x95CACC9: parse_token_trees_until_close_delim (tokentrees.rs:27)
    #8: 0x95CACC9: syntax::parse::lexer::tokentrees::<impl syntax::parse::lexer::StringReader<'a>>::parse_token_tree (tokentrees.rs:81)
  }
}
]]></programlisting>

<para>The first line indicates the node's position in the tree. The
<computeroutput>1.1</computeroutput> is a unique identifier for the node and
also says that it is the first child node <computeroutput>1</computeroutput>
(which is the root). The <computeroutput>/25</computeroutput> says that it is
one of 25 children, i.e. it has 24 siblings. The <computeroutput>(2
children)</computeroutput> says that this node node has two children of its
own.</para>

<para>Allocations are aggregated by their allocation stack trace. The
<computeroutput>Allocated at</computeroutput> section shows the allocation
stack trace that is shared by all the blocks covered by this node.</para>

<para>The <computeroutput>Total</computeroutput> line shows that this node
accounts for 4.02% of all bytes allocated during execution, and 7.72% of all
blocks. These percentages are useful for comparing the significance of
different nodes within a single profile; a PP that accounts for 10% of bytes
allocated is likely to be more interesting than one that accounts for
2%.</para>

<para>The <computeroutput>Total</computeroutput> line also shows allocation
rates, measured in bytes and blocks per million instructions. These rates are
useful for comparing the significance of nodes across profiles made with
different workloads.</para>

<para>Finally, the <computeroutput>Total</computeroutput> line shows the
average size and lifetimes of these blocks.</para>

<para>The <computeroutput>At t-gmax</computeroutput> line says shows that no
blocks from this PP were alive when the global heap peak occurred. In other
words, these blocks do not contribute at all to the global heap peak.</para>

<para>The <computeroutput>At t-end</computeroutput> line shows that no blocks
were from this PP were alive at shutdown. In other words, all those blocks were
explicitly freed before termination.</para>

<para>The <computeroutput>Reads</computeroutput> and
<computeroutput>Writes</computeroutput> lines show how many bytes were read 
within this PP's blocks, the fraction this represents of all heap reads, and
the read rate. Finally, it shows the read ratio, which is the number of reads
per byte. In this case the number is 0.29, which is quite low -- if no byte was
read twice, then only 29% of the allocated bytes, which means that at least 71%
of the bytes were never read! This suggests that the blocks are being
underutilized and might be worth optimizing.</para>

<para>The <computeroutput>Writes</computeroutput> lines is similar to the
<computeroutput>Reads</computeroutput> line. In this case, at most 38% of the
bytes are ever written, and at least 62% of the bytes were never written.
</para>

<para>The <computeroutput>Reads</computeroutput> and
<computeroutput>Writes</computeroutput> measurements suggest that the blocks
are being under-utilised and might be worth optimizing. Having said that, this
kind of under-utilisation is common in data structures that grow, such as
vectors and hash tables, and isn't always fixable. </para>

</sect3>


<sect3 id="dh-leaf-nodes"><title>Leaf Nodes</title>

<para>This is a leaf node:</para>

<programlisting><![CDATA[
PP 1.1.1.1/2 {
  Total:     31,460,928 bytes (2.32%, 1,565.9/Minstr) in 262,171 blocks (4.41%, 13.05/Minstr), avg size 120 bytes, avg lifetime 986,406,885.05 instrs (4.91% of program duration)
  Max:       16,779,136 bytes in 65,543 blocks, avg size 256 bytes
  At t-gmax: 0 bytes (0%) in 0 blocks (0%), avg size 0 bytes
  At t-end:  0 bytes (0%) in 0 blocks (0%), avg size 0 bytes
  Reads:     5,964,704 bytes (0.11%, 296.88/Minstr), 0.19/byte
  Writes:    10,487,200 bytes (0.51%, 521.98/Minstr), 0.33/byte
  Allocated at {
    ^1: 0x95CACC9: alloc (alloc.rs:72)
    ^2: 0x95CACC9: alloc (alloc.rs:148)
    ^3: 0x95CACC9: reserve_internal<syntax::tokenstream::TokenStream,alloc::alloc::Global> (raw_vec.rs:669)
    ^4: 0x95CACC9: reserve<syntax::tokenstream::TokenStream,alloc::alloc::Global> (raw_vec.rs:492)
    ^5: 0x95CACC9: reserve<syntax::tokenstream::TokenStream> (vec.rs:460)
    ^6: 0x95CACC9: push<syntax::tokenstream::TokenStream> (vec.rs:989)
    ^7: 0x95CACC9: parse_token_trees_until_close_delim (tokentrees.rs:27)
    ^8: 0x95CACC9: syntax::parse::lexer::tokentrees::<impl syntax::parse::lexer::StringReader<'a>>::parse_token_tree (tokentrees.rs:81)
    ^9: 0x95CAC39: parse_token_trees_until_close_delim (tokentrees.rs:26)
    ^10: 0x95CAC39: syntax::parse::lexer::tokentrees::<impl syntax::parse::lexer::StringReader<'a>>::parse_token_tree (tokentrees.rs:81)
    #11: 0x95CAC39: parse_token_trees_until_close_delim (tokentrees.rs:26)
    #12: 0x95CAC39: syntax::parse::lexer::tokentrees::<impl syntax::parse::lexer::StringReader<'a>>::parse_token_tree (tokentrees.rs:81)
  }
}
]]></programlisting>

<para>The <computeroutput>1.1.1.1/2</computeroutput> indicates that this node
is a great-grandchild of the root; is the first grandchild of the node in the
previous example; and has no children.</para>

<para>Leaf nodes contain an additional <computeroutput>Max</computeroutput>
line, indicating the peak memory use for the blocks covered by this PP. (This
peak may have occurred at a time other than
<computeroutput>t-gmax</computeroutput>.) In this case, 31,460,298 bytes were
allocated from this PP, but the maximum size alive at once was 16,779,136
bytes.</para>

<para>Stack frames that begin with a <computeroutput>^</computeroutput> rather
than a <computeroutput>#</computeroutput> are copied from ancestor nodes.
(In this example, the first 8 frames are identical to those from the node in
the previous example.) These frames could be found by tracing back through
ancestor nodes, but that can be annoying, which is why they are duplicated.
This also means that each node makes complete sense on its own.</para>

</sect3>


<sect3 id="dh-access-counts"><title>Access Counts</title>

<para>If all blocks covered by a PP node have the same size, an additional
<computeroutput>Accesses</computeroutput> field will be present. It indicates
how the reads and writes within these blocks were distributed. For
example:</para>

<programlisting><![CDATA[
Total:     8,388,672 bytes (0.62%, 417.53/Minstr) in 262,146 blocks (4.41%, 13.05/Minstr), avg size 32 bytes, avg lifetime 16,726,078,401.51 instrs (83.25% of program duration)
At t-gmax: 8,388,672 bytes (1.98%) in 262,146 blocks (16.64%), avg size 32 bytes
At t-end:  0 bytes (0%) in 0 blocks (0%), avg size 0 bytes
Reads:     9,109,682 bytes (0.17%, 453.41/Minstr), 1.09/byte
Writes:    7,340,088 bytes (0.36%, 365.34/Minstr), 0.88/byte
Accesses: {
  [  0]  65547 7 8 4 65529 〃 〃 〃 16 〃 〃 〃 12 〃 〃 〃 〃 〃 〃 〃 〃 〃 〃 〃 65542 〃 〃 〃 - - - - 
}
]]></programlisting>

<para>Every block covered by this PP was 32 bytes. Within all of those blocks,
byte 0 was accessed (read or written) 65,547 times, byte 1 was accessed 7
times, byte 2 was accessed 8 times, and so on.</para>

<para>The ditto symbol (<computeroutput>〃</computeroutput>) means "same access
count as the previous byte".</para>

<para>A dash (<computeroutput>-</computeroutput>) means "zero". (It is used
instead of <computeroutput>0</computeroutput> because it makes unaccessed
regions more easily identifiable.)</para>

<para>The infinity symbol (<computeroutput>∞</computeroutput>, not present in
this example) means "exceeded the maximum tracked count".</para>

<para>Block layout can often be inferred from counts. For example, these blocks
probably have four separate byte-sized fields, followed by a four-byte field,
and so on.</para>

<para>The size of the blocks that measure and display access counts is limited
to 1024 bytes. This is done to limit the performance overhead and also to keep
the size of the generated output reasonable. However, it is possible to override
this limit using client requests. The use-case for this is to first run DHAT
normally, and then identify any large blocks that you would like to further
investigate with access count histograms. The client request is declared in
<filename>dhat/dhat.h</filename> and is called <computeroutput>DHAT_HISTOGRAM_MEMORY</computeroutput>.
The macro should be placed immediately after the call to the allocator,
and use the pointer returned by the allocator.</para>

<programlisting><![CDATA[
// LargeStruct bigger than 1024 bytes
struct LargeStruct* ls = malloc(sizeof(struct LargeStruct));
DHAT_HISTOGRAM_MEMORY(ls);
]]></programlisting>

<para>The memory that can be profiled in this way with user requests
has a further upper limit of 25kbytes.  Be aware that the access counts
will all be set to zero. This means that the access counts will not
include any reads or writes performed during initialisation. An example where this
will happen are uses of C++ <computeroutput>new</computeroutput> with user-defined constructors.</para>

<para>Access counts can be useful for identifying data alignment holes or other
layout inefficiencies.</para>

</sect3>


<sect3 id="aggregate-nodes"><title>Aggregate Nodes</title>

<para>The PP tree is very large and many nodes represent tiny numbers of blocks
and bytes. Therefore, DHAT's viewer aggregates insignificant nodes like
this:</para>

<programlisting><![CDATA[
PP 1.14.2/2 {
  Total:     5,175 blocks (0.09%, 0.26/Minstr)
  Allocated at {
    [5 insignificant]
  }
}
]]></programlisting>

<para>Much of the detail is stripped away, leaving only basic measurements,
along with an indication of how many nodes were aggregated together (5 in this
case).</para>

</sect3>

</sect2>


<sect2 id="dh-output-footer"><title>The Output Footer</title>

<para>Below the PP tree is a line like this:</para>

<programlisting><![CDATA[
PP significance threshold: total >= 59,434.17 blocks (1%)
]]></programlisting>

<para>It shows the function used to determine if a PP node is significant. All
nodes that don't satisfy this function are aggregated. It is occasionally
useful if you don't understand why a PP node has been aggregated. The exact
threshold depends on the sort metric (see below).</para>

<para>Finally, the bottom of the page shows a legend that explains some of the
terms, abbreviations and symbols used in the output.</para>

</sect2>


<sect2 id="dh-sort-metrics"><title>Sort Metrics</title>

<para>The order in which sub-trees are sorted can be changed via the "Sort
metric" drop-down menu at the top of DHAT's viewer. Different sort metrics can
be useful for finding different things. Some sort metrics also incorporate some
filtering, so that only nodes meeting a particular criteria are shown.</para>

<!-- start of xi:include in the manpage -->
<variablelist>

  <varlistentry>
    <term>Total (bytes)</term>
    <listitem><para>The total number of bytes allocated during the execution.
    Highly useful for evaluating heap churn, though not quite as useful as
    "Total (blocks)".
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (blocks)</term>
    <listitem><para>The total number of blocks allocated during the execution.
    Highly useful for evaluating heap churn; reducing the number of calls to
    the allocator can significantly speed up a program. This is the default
    sort metric.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (blocks), tiny</term>
    <listitem><para>Like "Total (blocks)", but shows only very small blocks.
    Moderately useful, because such blocks are often easy to avoid allocating.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (blocks), short-lived</term>
    <listitem><para>Like "Total (blocks)", but shows only very short-lived
    blocks. Moderately useful, because such blocks are often easy to avoid
    allocating.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (bytes), zero reads or zero writes</term>
    <listitem><para>Like "Total (bytes)", but shows only blocks that are 
    never read or never written to (or both). Highly useful, because such
    blocks indicate poor use of memory and are often easy to avoid allocating.
    For example, sometimes a block is allocated and written to but then only
    read if a condition C is true; in that case, it may be possible to delay
    creating the block until condition C is true. Alternatively, sometimes
    blocks are created and never used; such blocks are trivial to remove.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (blocks), zero reads or zero writes</term>
    <listitem><para>Like "Total (bytes), zero reads or zero writes" but for
    blocks. Highly useful.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (bytes), low-access</term>
    <listitem><para>Like "Total (bytes)", but shows only blocks that have low
    numbers of reads or low numbers of writes (or both). Moderately useful,
    because such blocks indicate poor use of memory.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Total (blocks), low-access</term>
    <listitem><para>Like "Total (bytes), low-access", but for blocks.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>At t-gmax (bytes)</term>
    <listitem><para>This shows the breakdown of memory at the point of peak
    heap memory usage. Highly useful for reducing peak memory usage.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>At t-end (bytes)</term>
    <listitem><para>This shows the breakdown of memory at program termination.
    Highly useful for identifying process-lifetime leaks.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Reads (bytes)</term>
    <listitem><para>The number of bytes read within heap blocks. Occasionally
    useful.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Reads (bytes), high-access</term>
    <listitem><para>Like "Reads (bytes)", but only shows blocks with high read
    ratios. Occasionally useful for identifying hot areas of memory.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Writes (bytes)</term>
    <listitem><para>Like "Reads (bytes)", but for writes. Occasionally useful.
    </para></listitem>
  </varlistentry>

  <varlistentry>
    <term>Writes (bytes), high-access</term>
    <listitem><para>Like "Reads (bytes), high-access", but for writes.
    Occasionally useful.
    </para></listitem>
  </varlistentry>

</variablelist>

<para>The values within a node that represent the chosen sort metric are shown
in bold, so they stand out.</para>

<para>Here is part of a PP node found with "Total (blocks), tiny", showing
blocks with an average size of only 8.67 bytes:</para>

<programlisting><![CDATA[
Total:     3,407,848 bytes (0.25%, 169.62/Minstr) in 393,214 blocks (6.62%, 19.57/Minstr), avg size 8.67 bytes, avg lifetime 1,167,795,629.1 instrs (5.81% of program duration)
]]></programlisting>

<para>Here is part of a PP node found with "Total (blocks), short-lived",
showing blocks with an average lifetime of only 181.75 instructions:</para>

<programlisting><![CDATA[
Total:     23,068,584 bytes (1.7%, 1,148.19/Minstr) in 262,143 blocks (4.41%, 13.05/Minstr), avg size 88 bytes, avg lifetime 181.75 instrs (0% of program duration)
]]></programlisting>

<para>Here is an example of a PP identified with "Total (blocks), zero reads
or zero writes", showing blocks that are allocated but never touched:</para>

<programlisting><![CDATA[
Total:     7,339,920 bytes (0.54%, 365.33/Minstr) in 262,140 blocks (4.41%, 13.05/Minstr), avg size 28 bytes, avg lifetime 1,141,103,997.69 instrs (5.68% of program duration)
Max:       3,669,960 bytes in 131,070 blocks, avg size 28 bytes
At t-gmax: 3,336,400 bytes (0.79%) in 119,157 blocks (7.56%), avg size 28 bytes
At t-end:  0 bytes (0%) in 0 blocks (0%), avg size 0 bytes
Reads:     0 bytes (0%, 0/Minstr), 0/byte
Writes:    0 bytes (0%, 0/Minstr), 0/byte
]]></programlisting>

<para>All the blocks identified by these PPs are good candidates for
optimization.</para>

</sect2>

</sect1>


<sect1 id="dh-manual.realloc" xreflabel="Treatment of realloc">
<title>Treatment of realloc</title>

<para><computeroutput>realloc</computeroutput> is a tricky function and there
are several different ways that DHAT could handle it.</para>

<para>Imagine a <computeroutput>malloc(100)</computeroutput> call followed by
a <computeroutput>realloc(200)</computeroutput> call. This combination is
considered to add two to the total block count, and 300 bytes to the total
bytes count. (An alternative would be to only add one to the total block
count, and 200 bytes to the total bytes count, as if a single
<computeroutput>malloc(200)</computeroutput> call had occurred. While this
would be defensible from a semantic point of view, it is silly from an
operational point of view, because making two calls to allocator functions is
more expensive than one call, and DHAT is a profiler that aims to help with
runtime costs.)</para>

<para>Furthermore, the implicit copying of the 100 bytes is added to the reads
and writes counts. Without this, the read and write counts would be
under-measured and misleading.</para>

<para>However, DHAT only increases the current heap size by 100 bytes for this
combination, and does not change the current block count. (As opposed to
increasing the current heap size by 200 bytes and then decreasing it by 100
bytes.) As a result, it can only increase the global heap peak (if indeed,
this results in a new peak) by 100 bytes.</para>

<para>Finally, the program point assigned to the block allocated by the
<computeroutput>malloc(100)</computeroutput> call is retained once the block
is reallocated. Which means that all 300 bytes are attributed to that
program point, and no separate program point is created for the
<computeroutput>realloc(200)</computeroutput> call. This may be surprising,
but it has one large benefit.</para>

<para>Imagine some code that starts with an empty buffer, and then gradually
adds data to that buffer from numerous different points in the code,
reallocating the buffer each time it gets full. (E.g. code generation in a
compiler might work this way.) With the described approach, the first heap
block and all subsequent heap blocks are attributed to the same program point.
While this is something of a lie -- the first program point isn't actually
responsible for the other allocations -- it is arguably better than having the
program points spread around in a distribution that unpredictably depends on
whenever the reallocations were triggered.</para>

</sect1>


<sect1 id="dh-manual.copy-profiling" xreflabel="Copy profiling">
<title>Copy profiling</title>

<para>If DHAT is invoked with <option>--mode=copy</option>, instead of
profiling heap operations (allocations and deallocations), it profiles copy
operations, such as <computeroutput>memcpy</computeroutput>,
<computeroutput>memmove</computeroutput>,
<computeroutput>strcpy</computeroutput>, and
<computeroutput>bcopy</computeroutput>. This is sometimes useful.</para>

<para>Here is an example PP node from this mode:</para>

<programlisting><![CDATA[
PP 1.1.2/5 (4 children) {
  Total:     1,210,925 bytes (10.03%, 4,358.66/Minstr) in 112,717 blocks (35.2%, 405.72/Minstr), avg size 10.74 bytes
  Copied at {
    ^1: 0x4842524: memmove (vg_replace_strmem.c:1289)
    #2: 0x1F0A0D: copy_nonoverlapping<u8> (intrinsics.rs:1858)
    #3: 0x1F0A0D: copy_from_slice<u8> (mod.rs:2524)
    #4: 0x1F0A0D: spec_extend<u8> (vec.rs:2227)
    #5: 0x1F0A0D: extend_from_slice<u8> (vec.rs:1619)
    #6: 0x1F0A0D: push_str (string.rs:821)
    #7: 0x1F0A0D: write_str (string.rs:2418)
    #8: 0x1F0A0D: <&mut W as core::fmt::Write>::write_str (mod.rs:195)
  }
}
]]></programlisting>

<para>It is very similar to the PP nodes for heap profiling, but with less
information, because copy profiling doesn't involve any tracking of memory
regions with lifetimes.</para>

</sect1>


<sect1 id="dh-manual.ad-hoc-profiling" xreflabel="Ad hoc profiling">
<title>Ad hoc profiling</title>

<para>If DHAT is invoked with <option>--mode=ad-hoc</option>, instead of
profiling heap operations (allocations and deallocations), it profiles calls to
the <computeroutput>DHAT_AD_HOC_EVENT</computeroutput> client request, which is
declared in <filename>dhat/dhat.h</filename>.</para>

<para>Here is an example PP node from this mode:</para>

<programlisting><![CDATA[
PP 1.1.1.1/2 {
  Total:     30 units (17.65%, 115.97/Minstr) in 1 events (14.29%, 3.87/Minstr), avg size 30 units
  Occurred at {
    ^1: 0x109407: g (ad-hoc.c:4)
    ^2: 0x109425: f (ad-hoc.c:8)
    #3: 0x109497: main (ad-hoc.c:14)
  }
}
]]></programlisting>

<para>This kind of profiling is useful when you know a code path is hot but you
want to know more about it.</para>

<para>For example, you might want to know which callsites of a hot function
account for most of the calls. You could put a
<computeroutput>DHAT_AD_HOC_EVENT(1);</computeroutput> call at the start of
that function.</para>

<para>Alternatively, you might want to know the typical length of a vector in a
hot location. You could put a
<computeroutput>DHAT_AD_HOC_EVENT(len);</computeroutput> call at the
appropriate location, when <computeroutput>len</computeroutput> is the length
of the vector.</para>

</sect1>


<sect1 id="dh-manual.options" xreflabel="DHAT Command-line Options">
<title>DHAT Command-line Options</title>

<para>DHAT-specific command-line options are:</para>

<!-- start of xi:include in the manpage -->
<variablelist id="dh.opts.list">

  <varlistentry id="opt.dhat-out-file" xreflabel="--dhat-out-file">
    <term>
      <option><![CDATA[--dhat-out-file=<file> ]]></option>
    </term>
    <listitem>
      <para>Write the profile data to 
            <computeroutput>file</computeroutput> rather than to the default
            output file,
            <filename>dhat.out.&lt;pid&gt;</filename>. The
            <option>%p</option> and <option>%q</option> format specifiers
            can be used to embed the process ID and/or the contents of an
            environment variable in the name, as is the case for the core
            option <option><link linkend="opt.log-file">--log-file</link></option>.
      </para>
    </listitem>
  </varlistentry>

  <varlistentry id="opt.mode" xreflabel="--mode">
    <term>
      <option><![CDATA[--mode=<heap|copy|ad-hoc> [default: heap] ]]></option>
    </term>
    <listitem>
      <para>The profiling mode: heap profiling, copy profiling, or ad hoc
            profiling.
      </para>
    </listitem>
  </varlistentry>

</variablelist>

<para>Note that stacks by default have 12 frames. This may be more than
necessary, in which case the <option>--num-callers</option> flag can be used to
reduce the number, which may make DHAT run slightly faster.
</para>

<!-- end of xi:include in the manpage -->

</sect1>

</chapter>