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.

To enable sharing of derived files, use the &CacheDir; function in any &SConscript; file: CacheDir('/usr/local/build_cache') 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. 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. Actually, the MD5 signature is used as the name of the file in the shared cache directory in which the contents are stored. 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: % scons -Q cc -o hello.o -c hello.c cc -o hello hello.o % scons -Q -c Removed hello.o Removed hello % scons -Q Retrieved `hello.o' from cache Retrieved `hello' from cache 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 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.

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. If, however, you use the --cache-show option, &SCons; will print the command line that it would 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: % scons -Q cc -o hello.o -c hello.c cc -o hello hello.o % scons -Q -c Removed hello.o Removed hello % scons -Q --cache-show cc -o hello.o -c hello.c cc -o hello hello.o 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.

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: env = Environment() obj = env.Object('hello.c') env.Program('hello.c') CacheDir('cache') NoCache('hello.o') 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: % scons -Q cc -o hello.o -c hello.c cc -o hello hello.o % scons -Q -c Removed hello.o Removed hello % scons -Q cc -o hello.o -c hello.c Retrieved `hello' from cache

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. In these cases, you can specify the --cache-disable command-line option to tell &SCons; to not retrieve already-built files from the shared cache directory: % scons -Q cc -o hello.o -c hello.c cc -o hello hello.o % scons -Q -c Removed hello.o Removed hello % scons -Q Retrieved `hello.o' from cache Retrieved `hello' from cache % scons -Q -c Removed hello.o Removed hello % scons -Q --cache-disable cc -o hello.o -c hello.c cc -o hello hello.o

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. In this case, you can use the the --cache-force 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: % scons -Q --cache-disable cc -o hello.o -c hello.c cc -o hello hello.o % scons -Q -c Removed hello.o Removed hello % scons -Q --cache-disable cc -o hello.o -c hello.c cc -o hello hello.o % scons -Q --cache-force scons: `.' is up to date. % scons -Q scons: `.' is up to date. Notice how the above sample run demonstrates that the --cache-disable option avoids putting the built hello.o and hello files in the cache, but after using the --cache-force option, the files have been put in the cache for the next invocation to retrieve.

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: Program('prog', ['f1.c', 'f2.c', 'f3.c', 'f4.c', 'f5.c']) &SCons; will normally build the input object files on which the program depends in their normal, sorted order: % scons -Q 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 But if two such builds take place simultaneously, they may each look in the cache at nearly the same time and both decide that f1.o must be rebuilt and pushed into the shared cache directory, then both decide that f2.o must be rebuilt (and pushed into the shared cache directory), then both decide that f3.o 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. To alleviate such contention for the cache, you can use the --random command-line option to tell &SCons; to build dependencies in a random order: % scons -Q --random 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 Multiple builds using the --random 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. Note, of course, the --random 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. If you want to make sure dependencies will be built in a random order without having to specify the --random on very command line, you can use the &SetOption; function to set the random option within any &SConscript; file: 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'])