path: root/doc/user/caching.xml

<!--

  __COPYRIGHT__

  Permission is hereby granted, free of charge, to any person obtaining
  a copy of this software and associated documentation files (the
  "Software"), to deal in the Software without restriction, including
  without limitation the rights to use, copy, modify, merge, publish,
  distribute, sublicense, and/or sell copies of the Software, and to
  permit persons to whom the Software is furnished to do so, subject to
  the following conditions:

  The above copyright notice and this permission notice shall be included
  in all copies or substantial portions of the Software.

  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY
  KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE
  WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

-->

  <para>

  On multi-developer software projects,
  you can sometimes speed up every developer's builds a lot by
  allowing them to share the derived files that they build.
  &SCons; makes this easy, as well as reliable.

  </para>

  <section>
  <title>Specifying the Shared Cache Directory</title>

    <para>

    To enable sharing of derived files,
    use the &CacheDir; function
    in any &SConscript; file:

    </para>

    <programlisting>
       CacheDir('/usr/local/build_cache')
    </programlisting>

    <para>

    Note that the directory you specify must already exist
    and be readable and writable by all developers
    who will be sharing derived files.
    It should also be in some central location
    that all builds will be able to access.
    In environments where developers are using separate systems
    (like individual workstations) for builds,
    this directory would typically be
    on a shared or NFS-mounted file system.

    </para>

    <para>

    Here's what happens:
    When a build has a &CacheDir; specified,
    every time a file is built,
    it is stored in the shared cache directory
    along with its MD5 build signature.
      <footnote>
      <para>
      Actually, the MD5 signature is used as the name of the file
      in the shared cache directory in which the contents are stored.
      </para>
      </footnote>
    On subsequent builds,
    before an action is invoked to build a file,
    &SCons; will check the shared cache directory
    to see if a file with the exact same build
    signature already exists.
    If so, the derived file will not be built locally,
    but will be copied into the local build directory
    from the shared cache directory,
    like so:

    </para>

    <screen>
      % <userinput>scons -Q</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
      % <userinput>scons -Q -c</userinput>
      Removed hello.o
      Removed hello
      % <userinput>scons -Q</userinput>
      Retrieved `hello.o' from cache
      Retrieved `hello' from cache
    </screen>

    <para>

    Note that the &CacheDir; feature still calculates
    MD5 build sigantures for the shared cache file names
    even if you configure &SCons; to use timestamps
    to decide if files are up to date.
    (See the <xref linkend="chap-depends"></xref>
    chapter for information about the &Decider; function.)
    Consequently, using &CacheDir; may reduce or eliminate any
    potential performance improvements
    from using timestamps for up-to-date decisions.

    </para>

  </section>

  <section>
  <title>Keeping Build Output Consistent</title>

    <para>

    One potential drawback to using a shared cache
    is that the output printed by &SCons;
    can be inconsistent from invocation to invocation,
    because any given file may be rebuilt one time
    and retrieved from the shared cache the next time.
    This can make analyzing build output more difficult,
    especially for automated scripts that
    expect consistent output each time.

    </para>

    <para>

    If, however, you use the <literal>--cache-show</literal> option,
    &SCons; will print the command line that it
    <emphasis>would</emphasis> have executed
    to build the file,
    even when it is retrieving the file from the shared cache.
    This makes the build output consistent
    every time the build is run:

    </para>

    <screen>
      % <userinput>scons -Q</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
      % <userinput>scons -Q -c</userinput>
      Removed hello.o
      Removed hello
      % <userinput>scons -Q --cache-show</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
    </screen>

    <para>

    The trade-off, of course, is that you no longer
    know whether or not &SCons;
    has retrieved a derived file from cache
    or has rebuilt it locally.

    </para>

  </section>

  <section>
  <title>Not Using the Shared Cache for Specific Files</title>

    <para>

    You may want to disable caching for certain
    specific files in your configuration.
    For example, if you only want to put
    executable files in a central cache,
    but not the intermediate object files,
    you can use the &NoCache;
    function to specify that the
    object files should not be cached:

    </para>

    <programlisting>
       env = Environment()
       obj = env.Object('hello.c')
       env.Program('hello.c')
       CacheDir('cache')
       NoCache('hello.o')
    </programlisting>

    <para>

    Then when you run &scons; after cleaning
    the built targets,
    it will recompile the object file locally
    (since it doesn't exist in the shared cache directory),
    but still realize that the shared cache directory
    contains an up-to-date executable program
    that can be retrieved instead of re-linking:

    </para>

    <!--

    <scons_output example="ex1">
      <scons_output_command>scons -Q</scons_output_command>
      <scons_output_command>scons -Q -c</scons_output_command>
      <scons_output_command>scons -Q</scons_output_command>
    </scons_output>

    -->

    <screen>
      % <userinput>scons -Q</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
      % <userinput>scons -Q -c</userinput>
      Removed hello.o
      Removed hello
      % <userinput>scons -Q</userinput>
      cc -o hello.o -c hello.c
      Retrieved `hello' from cache
    </screen>

  </section>

  <section>
  <title>Disabling the Shared Cache</title>

    <para>

    Retrieving an already-built file
    from the shared cache
    is usually a significant time-savings
    over rebuilding the file,
    but how much of a savings
    (or even whether it saves time at all)
    can depend a great deal on your
    system or network configuration.
    For example, retrieving cached files
    from a busy server over a busy network
    might end up being slower than
    rebuilding the files locally.

    </para>

    <para>

    In these cases, you can specify
    the <literal>--cache-disable</literal>
    command-line option to tell &SCons;
    to not retrieve already-built files from the
    shared cache directory:

    </para>

    <screen>
      % <userinput>scons -Q</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
      % <userinput>scons -Q -c</userinput>
      Removed hello.o
      Removed hello
      % <userinput>scons -Q</userinput>
      Retrieved `hello.o' from cache
      Retrieved `hello' from cache
      % <userinput>scons -Q -c</userinput>
      Removed hello.o
      Removed hello
      % <userinput>scons -Q --cache-disable</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
    </screen>

  </section>

  <section>
  <title>Populating a Shared Cache With Already-Built Files</title>

    <para>

    Sometimes, you may have one or more derived files
    already built in your local build tree
    that you wish to make available to other people doing builds.
    For example, you may find it more effective to perform
    integration builds with the cache disabled
    (per the previous section)
    and only populate the shared cache directory
    with the built files after the integration build
    has completed successfully.
    This way, the cache will only get filled up
    with derived files that are part of a complete, successful build
    not with files that might be later overwritten
    while you debug integration problems.

    </para>

    <para>

    In this case, you can use the
    the <literal>--cache-force</literal> option
    to tell &SCons; to put all derived files in the cache,
    even if the files already exist in your local tree
    from having been built by a previous invocation:

    </para>

    <screen>
      % <userinput>scons -Q --cache-disable</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
      % <userinput>scons -Q -c</userinput>
      Removed hello.o
      Removed hello
      % <userinput>scons -Q --cache-disable</userinput>
      cc -o hello.o -c hello.c
      cc -o hello hello.o
      % <userinput>scons -Q --cache-force</userinput>
      scons: `.' is up to date.
      % <userinput>scons -Q</userinput>
      scons: `.' is up to date.
    </screen>

    <para>

    Notice how the above sample run
    demonstrates that the <literal>--cache-disable</literal>
    option avoids putting the built
    <filename>hello.o</filename>
    and
    <filename>hello</filename> files in the cache,
    but after using the <literal>--cache-force</literal> option,
    the files have been put in the cache
    for the next invocation to retrieve.

    </para>

  </section>

  <section>
  <title>Minimizing Cache Contention:  the <literal>--random</literal> Option</title>

    <para>

    If you allow multiple builds to update the
    shared cache directory simultaneously,
    two builds that occur at the same time
    can sometimes start "racing"
    with one another to build the same files
    in the same order.
    If, for example,
    you are linking multiple files into an executable program:

    </para>

    <programlisting>
       Program('prog',
               ['f1.c', 'f2.c', 'f3.c', 'f4.c', 'f5.c'])
    </programlisting>

    <para>

    &SCons; will normally build the input object files
    on which the program depends in their normal, sorted order:

    </para>

    <screen>
      % <userinput>scons -Q</userinput>
      cc -o f1.o -c f1.c
      cc -o f2.o -c f2.c
      cc -o f3.o -c f3.c
      cc -o f4.o -c f4.c
      cc -o f5.o -c f5.c
      cc -o prog f1.o f2.o f3.o f4.o f5.o
    </screen>

    <para>

    But if two such builds take place simultaneously,
    they may each look in the cache at nearly the same
    time and both decide that <filename>f1.o</filename>
    must be rebuilt and pushed into the shared cache directory,
    then both decide that <filename>f2.o</filename>
    must be rebuilt (and pushed into the shared cache directory),
    then both decide that <filename>f3.o</filename>
    must be rebuilt...
    This won't cause any actual build problems--both
    builds will succeed,
    generate correct output files,
    and populate the cache--but
    it does represent wasted effort.

    </para>

    <para>

    To alleviate such contention for the cache,
    you can use the <literal>--random</literal> command-line option
    to tell &SCons; to build dependencies
    in a random order:

    </para>

    <!--

    The following <screen> output was generated by this:

    <scons_output example="ex-random">
      <scons_output_command>scons -Q - -random</scons_output_command>
    </scons_output>

    We captured it directly here to guarantee a "random" order,
    guarding against the potential for - -random to happen
    to return things in the original sorted order.

    -->

    <screen>
      % <userinput>scons -Q --random</userinput>
      cc -o f3.o -c f3.c
      cc -o f1.o -c f1.c
      cc -o f5.o -c f5.c
      cc -o f2.o -c f2.c
      cc -o f4.o -c f4.c
      cc -o prog f1.o f2.o f3.o f4.o f5.o
    </screen>

    <para>

    Multiple builds using the <literal>--random</literal> option
    will usually build their dependencies in different,
    random orders,
    which minimizes the chances for a lot of
    contention for same-named files
    in the shared cache directory.
    Multiple simultaneous builds might still race to try to build
    the same target file on occasion,
    but long sequences of inefficient contention
    should be rare.

    </para>

    <para>

    Note, of course,
    the <literal>--random</literal> option
    will cause the output that &SCons; prints
    to be inconsistent from invocation to invocation,
    which may be an issue when
    trying to compare output from different build runs.

    </para>

    <para>

    If you want to make sure dependencies will be built
    in a random order without having to specify
    the <literal>--random</literal> on very command line,
    you can use the &SetOption; function to
    set the <literal>random</literal> option
    within any &SConscript; file:

    </para>

    <programlisting>
       Program('prog',
               ['f1.c', 'f2.c', 'f3.c', 'f4.c', 'f5.c'])

       SetOption('random', 1)
       Program('prog',
               ['f1.c', 'f2.c', 'f3.c', 'f4.c', 'f5.c'])
    </programlisting>

  </section>

  <!--

  <section>
  <title>Troubleshooting Shared Caching:  the &cache-debug; Option</title>

    <para>

    XXX describe the - - cache-debug option
    XXX maybe point to the troubleshooting appendix?

    </para>

  </section>

  -->

  <!--

  <section>

    <para>

    XXX describe CacheDir management:  monitoring, deleting, etc.

    </para>

  </section>

  -->