platform_build_soong/docs/perf.md

Build Performance

Debugging Build Performance

Tracing

soong_ui has tracing built in, so that every build execution's trace can be viewed. Just open $OUT_DIR/build.trace.gz in Chrome's chrome://tracing, or with catapult's trace viewer. The last few traces are stored in build.trace.#.gz (larger numbers are older). The associated logs are stored in soong.#.log and verbose.#.log.gz.

Soong

Soong can be traced and profiled using the standard Go tools. It understands the -cpuprofile, -trace, and -memprofile command line arguments, but we don't currently have an easy way to enable them in the context of a full build.

Kati

In general, the slow path of reading Android.mk files isn't particularly performance sensitive, since it doesn't need to happen on every build. It is important for the fast-path (detecting whether it needs to regenerate the ninja file) to be fast however. And it shouldn't hit the slow path too often -- so don't rely on output of a $(shell) command that includes the current timestamp, or read a file that's going to change on every build.

Regen check is slow

In most cases, we've found that the fast-path is slow because all of the $(shell) commands need to be re-executed to determine if their output changed. The $OUT_DIR/verbose.log.gz contains statistics from the regen check:

verbose: *kati*: regen check time: 0.754030
verbose: *kati*: glob time (regen): 0.545859 / 43840
verbose: *kati*: shell time (regen): 0.278095 / 66 (59 unique)
verbose: *kati*:   0.012 / 1 mkdir -p out/target/product/generic && echo Android/aosp_arm/generic:R/AOSP.MASTER/$(date -d @$(cat out/build_date.txt) +%m%d%H%M):eng/test-keys >out/target/product/generic/build_fingerprint.txt && grep " " out/target/product/generic/build_fingerprint.txt
verbose: *kati*:   0.010 / 1 echo 'com.android.launcher3.config.FlagOverrideSampleTest com.android.launcher3.logging.FileLogTest com.android.launcher3.model.AddWorkspaceItemsTaskTest com.android.launcher3.model.CacheDataUpdatedTaskTest com.android.launcher3.model.DbDowngradeHelperTest com.android.launcher3.model.GridBackupTableTest com.android.launcher3.model.GridSizeMigrationTaskTest com.android.launcher3.model.PackageInstallStateChangedTaskTest com.android.launcher3.popup.PopupPopulatorTest com.android.launcher3.util.GridOccupancyTest com.android.launcher3.util.IntSetTest' | tr ' ' '\n' | cat
verbose: *kati*:   0.010 / 1 cd cts/tests/framework/base/windowmanager ; find -L  * -name "Components.java" -and -not -name ".*"
verbose: *kati*:   0.010 / 1 git -C test/framework/build log -s -n 1 --format="%cd" --date=format:"%Y%m%d_%H%M%S" 2>/dev/null
verbose: *kati*:   0.009 / 2 cd development/samples/ShortcutDemo/publisher ; find -L  ../common/src -name "*.java" -and -not -name ".*"
verbose: *kati*:   0.009 / 2 cd development/samples/ShortcutDemo/launcher ; find -L  ../common/src -name "*.java" -and -not -name ".*"
verbose: *kati*:   0.009 / 1 if ! cmp -s out/target/product/generic/obj/CONFIG/kati_packaging/dist.mk.tmp out/target/product/generic/obj/CONFIG/kati_packaging/dist.mk; then mv out/target/product/generic/obj/CONFIG/kati_packaging/dist.mk.tmp out/target/product/generic/obj/CONFIG/kati_packaging/dist.mk; else rm out/target/product/generic/obj/CONFIG/kati_packaging/dist.mk.tmp; fi
verbose: *kati*:   0.008 / 1 mkdir -p out/target/product/generic && echo R/AOSP.MASTER/$(cat out/build_number.txt):eng/test-keys >out/target/product/generic/build_thumbprint.txt && grep " " out/target/product/generic/build_thumbprint.txt
verbose: *kati*:   0.007 / 1 echo 'com.android.customization.model.clock.BaseClockManagerTest com.android.customization.model.clock.ClockManagerTest com.android.customization.model.grid.GridOptionsManagerTest com.android.customization.model.theme.ThemeManagerTest' | tr ' ' '\n' | cat
verbose: *kati*:   0.007 / 1 uname -sm
verbose: *kati*: stat time (regen): 0.361907 / 1241

In this case, the total time spent checking was 0.75 seconds, even though the other "(regen)" numbers add up to more than that (some parts are parallelized where possible). Often times, the biggest contributor is the $(shell) times -- in this case, 66 calls took 0.27s. The top 10 longest shell functions are printed.

All the longest commands in this case are all variants of a call to find, but this is where using pure make functions instead of calling out to the shell can make a performance impact -- many calls to check if 26 > 20 can add up. We've added some basic math functions in math.mk to help some common use cases that used to be rather expensive when they were used too often.

There are some optimizations in place for find commands -- if Kati can understand the find command, the built-in find emulator can turn some of them into glob or stat checks (falling back to calling find if one of those imply that the output may change). Many of the common macros produce find commands that Kati can understand, but if you're writing your own, you may want to experiment with other options if they're showing up in this list. For example, if this was significantly more expensive (either in runtime, or was called often):

.../kati.go:127: *kati*:  0.015 cd libcore && (find luni/src/test/java -name "*.java" 2> /dev/null) | grep -v -f java_tests_blacklist

It may be more efficient to move the grep into make, so that the find portion can be rewritten and cached:

$(filter-out $(file <$(LOCAL_PATH)/java_tests_blacklist),$(call all-java-files-under,luni/src/test/java))

Others can be simplified by just switching to an equivalent find command that Kati understands:

.../kati.go:127: *kati*:  0.217 find device vendor -type f -name \*.pk8 -o -name verifiedboot\* -o -name \*.x509.pem -o -name oem\*.prop | sort

By adding the implicit -a and moving the | sort to Make, this can now be cached by Kati:

$(sort $(shell find device vendor -type -f -a -name \*.pk8 -o -name verifiedboot\* -o -name \*.x509.pem -o -name oem\*.prop))

Kati has now learned about the implicit -a, so this particular change is no longer necessary, but the basic concept holds.

Kati regens too often

Kati prints out what triggered the slow path to be taken -- this can be a changed file, a changed environment variable, or different output from a $(shell) command:

out/soong/Android-aosp_arm.mk was modified, regenerating...

The state is stored in $OUT_DIR/.kati_stamp* files, and can be (partially) read with the ckati_stamp_dump tool in prebuilts/build-tools. More debugging is available when ckati is run with --regen_debug, but that can be a lot of data to understand.

Debugging the slow path

Kati will now dump out information about which Makefiles took the most time to execute. This is also in the verbose.log.gz file:

verbose: *kati*: included makefiles: 73.640833 / 232810 (1066 unique)
verbose: *kati*:  18.389 /     1 out/soong/Android-aosp_arm.mk
verbose: *kati*:  13.137 / 20144 build/make/core/soong_cc_prebuilt.mk
verbose: *kati*:  11.743 / 27666 build/make/core/base_rules.mk
verbose: *kati*:   2.289 /     1 art/Android.mk
verbose: *kati*:   2.054 /     1 art/build/Android.cpplint.mk
verbose: *kati*:   1.955 / 28269 build/make/core/clear_vars.mk
verbose: *kati*:   1.795 /   283 build/make/core/package.mk
verbose: *kati*:   1.790 /   283 build/make/core/package_internal.mk
verbose: *kati*:   1.757 / 17382 build/make/core/link_type.mk
verbose: *kati*:   1.078 /   297 build/make/core/aapt2.mk

This shows that soong_cc_prebuilt.mk was included 20144 times, for a total time spent of 13.137 secounds. While Android-aosp_arm.mk was only included once, and took 18.389 seconds. In this case, Android-aosp_arm.mk is the only file that includes soong_cc_prebuilt.mk, so we can safely assume that 13 of the 18 seconds in Android-aosp_arm.mk was actually spent within soong_cc_prebuilt.mk (or something that it included, like base_rules.mk).

By default this only includes the top 10 entries, but you can ask for the stats for any makefile to be printed with $(KATI_profile_makefile):

$(KATI_profile_makefile build/make/core/product.mk)

With these primitives, it's possible to get the timing information for small chunks, or even single lines, of a makefile. Just move the lines you want to measure into a new makefile, and replace their use with an include of the new makefile. It's possible to analyze where the time is being spent by doing a binary search using this method, but you do need to be careful not to split conditionals across two files (the ifeq/else/endif must all be in the same file).

Ninja

Understanding why something rebuilt

Add NINJA_ARGS="-d explain" to your environment before a build, this will cause ninja to print out explanations on why actions were taken. Start reading from the beginning, as this much data can be hard to read:

$ cd art
$ mma
$ touch runtime/jit/profile_compilation_info.h
$ NINJA_ARGS="-d explain" mma
...
ninja explain: output out/soong/.intermediates/art/tools/cpp-define-generator/cpp-define-generator-data/linux_glibc_x86_64/obj/art/tools/cpp-define-generator/main.o older than most recent input art/runtime/jit/profile_compilation_info.h (
1516683538 vs 1516685188)
ninja explain: out/soong/.intermediates/art/tools/cpp-define-generator/cpp-define-generator-data/linux_glibc_x86_64/obj/art/tools/cpp-define-generator/main.o is dirty
ninja explain: out/soong/.intermediates/art/tools/cpp-define-generator/cpp-define-generator-data/linux_glibc_x86_64/cpp-define-generator-data is dirty
ninja explain: out/soong/host/linux-x86/bin/cpp-define-generator-data is dirty
ninja explain: out/soong/.intermediates/art/tools/cpp-define-generator/cpp-define-generator-asm-support/gen/asm_support_gen.h is dirty
ninja explain: out/soong/.intermediates/art/cmdline/art_cmdline_tests/android_arm_armv7-a_core_cmdline_parser_test/obj/art/cmdline/cmdline_parser_test.o is dirty
...

In this case, art/cmdline/cmdline_parser_test.o was rebuilt because it uses asm_support_gen.h, which was generated by cpp-define-generator-data, which uses profile_compilation_info.h.

You'll likely need to cross-reference this data against the build graph in the various .ninja files. The files are (mostly) human-readable, but a (slow) web interface can be used by running NINJA_ARGS="-t browse <target>" m.

Builds take a long time

If the long part in the trace view of a build is a relatively solid block, then the performance is probably more related to how much time the actual build commands are taking than having extra dependencies, or slowdowns in soong/kati/ninja themselves.

Beyond looking at visible outliers in the trace view, we don't have any tooling to help in this area yet. It's possible to aggregate some of the raw data together, but since our builds are heavily parallelized, it's particularly easy for build commands to impact unrelated build commands. This is an area we'd like to improve -- we expect keeping track of user/system time per-action would provide more reliable data, but tracking some full-system data (memory/swap use, disk bandwidth, etc) may also be necessary.

Known Issues

Common

<= Android 10 (Q): mm

Soong always loads the entire module graph, so as modules convert from Make to Soong, mm is becoming closer to mma. This produces more correct builds, but does slow down builds, as we need to verify/produce/load a larger build graph.

As of Android Q, loading large build graphs is fast, and in Android R, mm is now an alias of mma.

Android 8.1 (Oreo MR1)

In some cases, a tree would get into a state where Soong would be run twice on every incremental build, even if there was nothing to do. This was fixed in master with these changes, but they were too significant to backport at the time. And while they fix this particular issue, they appear to cause ninja to spend more time during every build loading the .ninja_log / .ninja_deps files, especially as they become larger.

A workaround to get out of this state is to remove the build.ninja entry from $OUT_DIR/.ninja_log:

sed -i "/\/build.ninja/d" $(get_build_var OUT_DIR)/.ninja_log