Issue metadata
Sign in to add a comment
|
clang tot linux bots mysteriously faster (while still on gyp) |
||||||||||||||||||||
Issue descriptionhttps://build.chromium.org/p/chromium.fyi/builders/ClangToTLinuxASan/builds/3943/steps/gclient%20runhooks/logs/stdio 'asan=1 clang=1 clang_use_chrome_plugins=0 component=static_library lsan=1 target_arch=x64 test_isolation_mode=prepare' https://build.chromium.org/p/chromium.fyi/builders/ClangToTLinuxASan/builds/3962/steps/generate_build_files/logs/stdio GYP_DEFINES='test_isolation_mode=prepare clang_use_chrome_plugins=0 clang=1 asan=1 lsan=1 component=static_library gomadir=/b/build/slave/cache/cipd/goma' Note that target_arch=x64 disappeared. x64 is the default in gn, but not in gyp. Not fixing this myself since this is potentially a systemic issue with the mb conversion process.
,
Apr 26 2016
Hm, that's a good point. However, something did happen, see https://build.chromium.org/p/chromium.fyi/stats/ClangToTLinuxASan You can see where mb got turned on by when "runhooks" time drops, and build time dropped around the same time. The bot still uses gyp and builds the same thing, so it shouldn't suddenly get faster. I ssh'd in and checked that the ninja files that gyp generates do pass -m64 (they do), so it's still a 64-bit build.
,
Apr 26 2016
The last pre-MB build is build 3958, which clocks in at 2:48:40. Build 3959 fails to compile; build 3960 takes 2:43:54. I don't see any obvious differences in how we're generating build flags, so my guess is that something actually landed that makes the compile take less time. Though, oddly, we're building more things (31139 edges in ninja vs. 31803). It's just a guess, though.
,
Apr 26 2016
Oh hm, I guess one thing that's new is that mb is supposed to set LLVM_FORCE_HEAD_REVISION, and if it doesn't then build/common.gypi won't define CR_CLANG_REVISION to the trunk revision but to the pinned one, and the bot won't do clobber builds. ...but no, it looks like base/base.ninja on the bot does have -DCR_CLANG_REVISION=267552 which looks right.
,
Apr 26 2016
Last slow: https://build.chromium.org/p/chromium.fyi/builders/ClangToTLinuxASan/builds/3925 First fast: https://build.chromium.org/p/chromium.fyi/builders/ClangToTLinuxASan/builds/3928 This happened before we switched to mb, so this is due to something else.
,
Apr 26 2016
https://build.chromium.org/p/chromium.fyi/builders/ClangToTLinuxASan/builds/3927 has the protoc update. I think I'd buy that this might improve build times by 10% or so, but then why did that happen on Linux only? Anyhow, something happened that made builds faster (hooray). If we can figure out what, maybe we can make that happen on other bots too.
,
Apr 26 2016
I wanted to compare build traces, but that's apparently broke atm, filed bug 606974 for that.
,
Apr 27 2016
Slower build: http://chromium-build-stats.appspot.com/ninja_log/2016/04/16/slave4-c1/ninja_log.slave4-c1.chrome-bot.20160416-164504.13675.gz Faster build: http://chromium-build-stats.appspot.com/ninja_log/2016/04/21/slave4-c1/ninja_log.slave4-c1.chrome-bot.20160421-123954.22271.gz Looks like everything is a bit faster (look at the table; it's sorted by duration.) It's not just compiles, links are faster too, so probably not a change in clang.
,
Apr 27 2016
I guess we'll never know. |
|||||||||||||||||||||
►
Sign in to add a comment |
|||||||||||||||||||||
Comment 1 by dpranke@chromium.org
, Apr 26 2016