Skip to content

Latest commit

 

History

History
168 lines (115 loc) · 5.58 KB

linux_faster_builds.md

File metadata and controls

168 lines (115 loc) · 5.58 KB

Tips for improving build speed on Linux

This list is sorted such that the largest speedup is first; see Linux build instructions for context and Faster Builds for non-Linux-specific techniques.

[TOC]

Use goma

If you work at Google, you can use goma for distributed builds; this is similar to distcc. See go/ma for documentation.

Even without goma, you can do distributed builds with distcc (if you have access to other machines), or a parallel build locally if have multiple cores.

Whether using goma, distcc, or parallel building, you can specify the number of build processes with -jX where X is the number of processes to start.

Use Icecc

Icecc is the distributed compiler with a central scheduler to share build load. Currently, many external contributors use it. e.g. Intel, Opera, Samsung.

When you use Icecc, you need to set some gyp variables.

linux_use_bundled_binutils=0**

-B option is not supported. relevant commit

linux_use_debug_fission=0

debug fission is not supported. bug

clang=0

Icecc doesn't support clang yet.

use_sysroot=0

Icecc doesn't work with sysroot.

linux_use_bundled_gold=0

Using the system linker is necessary when using glibc 2.21 or newer. See related bug.

Build only specific targets

If you specify just the target(s) you want built, the build will only walk that portion of the dependency graph:

cd $CHROMIUM_ROOT/src
ninja -C out/Debug base_unittests

Linking

Dynamically link

We normally statically link everything into one final executable, which produces enormous (nearly 1gb in debug mode) files. If you dynamically link, you save a lot of time linking for a bit of time during startup, which is fine especially when you're in an edit/compile/test cycle.

Run gyp with the -Dcomponent=shared_library flag to put it in this configuration. (Or set those flags via the GYP_DEFINES environment variable.)

e.g.

build/gyp_chromium -D component=shared_library
ninja -C out/Debug chrome

See the component build page for more information.

Linking using gold

The experimental "gold" linker is much faster than the standard BFD linker.

On some systems (including Debian experimental, Ubuntu Karmic and beyond), there exists a binutils-gold package. Do not install this version! Having gold as the default linker is known to break kernel / kernel module builds.

The Chrome tree now includes a binary of gold compiled for x64 Linux. It is used by default on those systems.

On other systems, to safely install gold, make sure the final binary is named ld and then set CC/CXX appropriately, e.g. export CC="gcc -B/usr/local/gold/bin" and similarly for CXX. Alternatively, you can add /usr/local/gold/bin to your PATH in front of /usr/bin.

WebKit

Build WebKit without debug symbols

WebKit is about half our weight in terms of debug symbols. (Lots of templates!) If you're working on UI bits where you don't care to trace into WebKit you can cut down the size and slowness of debug builds significantly by building WebKit without debug symbols.

Set the gyp variable remove_webcore_debug_symbols=1, either via the GYP_DEFINES environment variable, the -D flag to gyp, or by adding the following to ~/.gyp/include.gypi:

{
  'variables': {
    'remove_webcore_debug_symbols': 1,
  },
}

Tune ccache for multiple working directories

(Ignore this if you use goma.)

Increase your ccache hit rate by setting CCACHE_BASEDIR to a parent directory that the working directories all have in common (e.g., /home/yourusername/development). Consider using CCACHE_SLOPPINESS=include_file_mtime (since if you are using multiple working directories, header times in svn sync'ed portions of your trees will be different - see the ccache troubleshooting section for additional information). If you use symbolic links from your home directory to get to the local physical disk directory where you keep those working development directories, consider putting

alias cd="cd -P"

in your .bashrc so that $PWD or cwd always refers to a physical, not logical directory (and make sure CCACHE_BASEDIR also refers to a physical parent).

If you tune ccache correctly, a second working directory that uses a branch tracking trunk and is up-to-date with trunk and was gclient sync'ed at about the same time should build chrome in about 1/3 the time, and the cache misses as reported by ccache -s should barely increase.

This is especially useful if you use git-new-workdir and keep multiple local working directories going at once.

Using tmpfs

You can use tmpfs for the build output to reduce the amount of disk writes required. I.e. mount tmpfs to the output directory where the build output goes:

As root:

mount -t tmpfs -o size=20G,nr_inodes=40k,mode=1777 tmpfs /path/to/out

*** note Caveat: You need to have enough RAM + swap to back the tmpfs. For a full debug build, you will need about 20 GB. Less for just building the chrome target or for a release build.

Quick and dirty benchmark numbers on a HP Z600 (Intel core i7, 16 cores hyperthreaded, 12 GB RAM)

  • With tmpfs:
    • 12m:20s
  • Without tmpfs
    • 15m:40s