A tool for gathering profiling and benchmarking information for Gradle builds
A tool to automate the gathering of profiling and benchmarking information for Gradle builds.
Profiling information can be captured using several different tools:
SDKMAN! is a tool for managing parallel versions of multiple Software Development Kits on most Unix-based systems.
> sdk install gradleprofiler
> gradle-profiler --benchmark help
Homebrew is the easiest and most flexible way to install the UNIX tools Apple didn’t include with macOS.
> brew install gradle-profiler
> gradle-profiler --benchmark help
Binaries are available and linked from the releases page.
First, build and install the gradle-profiler app using:
> ./gradlew installDist
This will install the executable into ./build/install/gradle-profiler/bin. The examples below assume that you add this location to your PATH or create a gradle-profiler alias for it.
NOTE: You have to use Java 11 or later to build this project.
Benchmarking simply records the time it takes to execute your build several times and calculates a mean and standard error for it.
It has zero impact on the execution time, so it is ideal for making before/after comparisons for new Gradle versions or changes to your build.
Run the gradle-profiler app using:
> gradle-profiler --benchmark --project-dir <root-dir-of-build> <task>...
Where <root-dir-of-build> is the directory containing the build to be benchmarked, and <task> is the name of the task to run,
exactly as you would use for the gradle command.
Results will be written to a file called profile-out/benchmark.html and profile-out/benchmark.csv.
When the profiler runs the build, it will use the tasks you specified. The profiler will use the default
Gradle version, Java installation and JVM args that have been specified for your build, if any.
This generally works the same way as if you were using the Gradle wrapper. For example, the profiler will use the values
from your Gradle wrapper properties file, if present, to determine which Gradle version to run.
You can use the --gradle-version option to specify a Gradle version or installation to use to benchmark the build.
You can specify multiple versions and each of these is used to benchmark the build, allowing you to compare the behaviour of several different Gradle versions.
You can also use the --measure-config-time option to measure some additional details about configuration time.
You can use --measure-build-op together with the fully qualified class name of the enveloping type of the Details interface to benchmark cumulative build operation time.
For example, for Gradle 5.x there is a org.gradle.api.internal.tasks.SnapshotTaskInputsBuildOperationType which can be used to capture snapshotting time.
The time recorded is cumulative time, so the wall clock time spent on executing the measured build operations is probably smaller.
If the build operation does not exists in a benchmarked version of Gradle, it is gracefully ignored.
In the resulting reports it will show up with 0 time.
If multiple versions are tested, then Gradle profiler determines whether there is an statistically significant difference in the run times by using a Mann-Whitney U-Test.
The result files contain the confidence if a sample has a different performance behavior - i.e. it is faster or slower - than the baseline.
Profiling allows you to get deeper insight into the performance of your build.
To run the gradle-profiler app to profile a build use:
> gradle-profiler --profile <name-of-profiler> --project-dir <root-dir-of-build> <task>...
The app will run the build several times to warm up a daemon, then enable the profiler and run the build.
Once complete, the results are available under profile-out/.
If you use Async profiler or JFR for profiling, Gradle profiler will also create flame graphs for each scenario.
If you profile multiple scenarios or multiple versions, then Gradle profiler will create differential flame graphs as well.
Gradle build scans are a powerful tool to investigate the structure of your build and quickly find bottlenecks.
You can use the timeline view to see which tasks ran, how long they took, whether they were cached, how well your build parallelized etc.
The performance tab will show you details about configuration time and other hints on how to make your build faster.
In order to create a build scan of your build, use --profile buildscan. The build scan URL is reported on the console and is also available in profile-out/profile.log.
Async profiler provides low-overhead CPU, allocation and perf event sampling on Linux and MacOS.
It also correctly handles native method calls, making it preferable to JFR on these operating systems.
You can use async profiler to profile a Gradle build using --profile async-profiler. By default, this will profile CPU usage, with some reasonable default settings. These settings can be configured using various command-line options, listed below.
Alternatively, you can also use --profile async-profiler-heap to profile heap allocations, with some reasonable default settings.
Finally, you can also use --profile async-profiler-all to profile cpu, heap allocations, and locks with some reasonable default settings.
By default, an Async profiler release will be downloaded from Github and installed, if not already available.
The output are flame and icicle graphs which show you the call tree and hotspots of your code.
The following options are supported and closely mimic the options of Async profiler. Have a look at its readme to find out more about each option:
--async-profiler-event: The event to sample, e.g. cpu, wall, lock or alloc. Defaults to cpu. Multiple events can be profiled by using this parameter multiple times.--async-profiler-count: The count to use when aggregating event data. Either samples or total. total is especially useful for allocation profiling. Defaults to samples. Corresponds to the --samples and --total command line options for Async profiler.--async-profiler-interval: The sampling interval in ns, defaults to 10_000_000 (10 ms).--async-profiler-alloc-interval: The sampling interval in bytes for allocation profiling, defaults to 10 bytes. Corresponds to the --alloc command line option for Async profiler.--async-profiler-lock-threshold: lock profiling threshold in nanoseconds, defaults to 250 microseconds. Corresponds to the --lock command line option for Async profiler.--async-profiler-stackdepth: The maximum stack depth. Lower this if profiles with deep recursion get too large. Defaults to 2048.--async-profiler-system-threads: Whether to show system threads like GC and JIT compilation in the profile. Usually makes them harder to read, but can be useful if you suspect problems in that area. Defaults to false.You can also use either the ASYNC_PROFILER_HOME environment variable or the --async-profiler-home command line option to point to the Async profiler install directory.
JProfiler is a powerful commercial profiler, which provides both sampling and instrumentation capabilities.
You can tailor its settings in the JProfiler UI and then instruct the Gradle profiler to use these settings for full control
over what you want to investigate. For instance, you could split calls to a dependency resolution rule by argument to
find out if the rule is slow for a specific dependency.
In order to work with JProfiler, use the --profile jprofiler option.
This will use JProfiler’s CPU sampling by default. JProfiler supports several other options:
--jprofiler-config sampling-all (by default only packages containing the word gradle are sampled)--jprofiler-config instrumentation--jprofiler-alloc--jprofiler-monitors--jprofiler-probes:<probe ids, separated by comma> (e.g. --jprofiler-probes builtin.FileProbe)--jprofiler-heapdump--jprofiler-session <sessionId>--jprofiler-home /path/to/jprofilerYourKit is a powerful commercial profiler, which provides both sampling and instrumentation capabilities.
Its integration in the Gradle profiler is currently limited, e.g. support for probes and other custom settings
is missing. If you are using YourKit and would like to see better support, pull requests are welcome.
In order to work with YourKit, make sure the YOURKIT_HOME environment variable is set and then use the --profile yourkit option. This will use YourKit’s CPU sampling instrumentation by default.
You can switch to CPU tracing using the --profile yourkit-tracing option.
You can switch to memory allocation profiling by using the --profile yourkit-heap option.
All probes are disabled when using sampling or memory allocation profiling.
JFR provides low overhead CPU, allocation, IO wait and lock profiling and runs on all major operating systems.
It is available on Oracle JDK since Java 7 and on OpenJDK since Java 11 (make sure you have at least 11.0.3).
To our knowledge, it is the only low-overhead allocation profiler for Windows.
However, be aware of its shortcomings, e.g. it will not sample native method calls, so you will get misleading CPU results if your code does a lot of system calls (like reading files).
You will get both the JFR file and flame graph visualizations of the data, which are much easier to understand than the Java Mission Control UI.
In order to profile with JFR, add the --profile jfr option.
You can change the profiler settings using --jfr-settings, specifying either the path to a .jfc file or the name of a built-in template like profile.
To capture a heap dump at the end of each measured build, add the --profile heap-dump option. You can use this with other --profile options.
Chrome traces are a low-level event dump (e.g. projects being evaluated, tasks being run etc.).
They are useful when you can’t create a build scan, but need to look at the overall structure of a build.
It also displays CPU load, memory usage and GC activity. Using chrome-trace requires Gradle 3.3+.
Add the --profile chrome-trace option and open the result in Google Chrome in chrome://tracing.
--project-dir: Directory containing the build to run (required).--benchmark: Benchmark the build. Runs the builds more times and writes the results to a CSV file.--profile <profiler>: Profile the build using the specified profiler. See above for details on each profiler.--output-dir <dir>: Directory to write results to. Default value is profile-out. If profile-out directory already exists, it tries to find a profile-out-<index> directory, that does not exist.--warmups: Specifies the number of warm-up builds to run for each scenario. Defaults to 2 for profiling, 6 for benchmarking, and 1 when not using a warm daemon.--iterations: Specifies the number of builds to run for each scenario. Defaults to 1 for profiling, 10 for benchmarking.--bazel: Benchmark scenarios using Bazel instead of Gradle. By default, only Gradle scenarios are run. You cannot profile a Bazel build using this tool.--buck: Benchmark scenarios using Buck instead of Gradle. By default, only Gradle scenarios are run. You cannot profile a Buck build using this tool.--maven: Benchmark scenarios using Maven instead of Gradle. By default, only Gradle scenarios are run. You cannot profile a Maven build using this tool.The following command line options only apply when measuring Gradle builds:
--gradle-user-home: The Gradle user home. Defaults to <project-dir>/gradle-user-home to isolate performance tests from your other builds.--gradle-version <version>: Specifies a Gradle version or installation to use to run the builds, overriding the default for the build. You can specify multiple versions by using this option once for each version.--no-daemon: Uses the gradle command-line client with the --no-daemon option to run the builds. The default is to use the Gradle tooling API and Gradle daemon.--cold-daemon: Use a cold daemon (one that has just started) rather than a warm daemon (one that has already run some builds). The default is to use a warm daemon.--cli: Uses the gradle command-line client to run the builds. The default is to use the Gradle tooling API and Gradle daemon.--measure-build-op: Additionally measure the cumulative time spent in the given build operation. Only supported for Gradle 6.1 and later.--measure-config-time: Measure some additional details about configuration time. Only supported for Gradle 6.1 and later.--measure-gc: Measure the garbage collection time. Only supported for Gradle 6.1 and later.--measure-local-build-cache: Measure the size of the local build cache.-D<key>=<value>: Defines a system property when running the build, overriding the default for the build.--studio-install-dir: The Android Studio installation directory. Required when measuring Android Studio sync.--studio-sandbox-dir: The Android Studio sandbox directory. It’s recommended to use it since it isolates the Android Studio process from your other Android Studio processes. By default, this will be set to <output-dir>/studio-sandbox. If you want Android Studio to keep old data (e.g. indexes) you should set and reuse your own folder.--no-studio-sandbox: Do not use the Android Studio sandbox but use the default Android Studio folders for the Android Studio data.--no-diffs: Do not generate differential flame graphs.A scenario file can be provided to define more complex scenarios to benchmark or profile. Use the --scenario-file option to provide this. The scenario file is defined in Typesafe config format.
The scenario file defines one or more scenarios. You can select which scenarios to run by specifying its name on the command-line when running gradle-profiler, e.g.
> gradle-profiler --benchmark --scenario-file performance.scenarios clean_build
Here is an example:
# Can specify scenarios to use when none are specified on the command line
default-scenarios = ["assemble"]
# Scenarios are run in alphabetical order
assemble {
# Show a slightly more human-readable title in reports
title = "Assemble"
# Run the 'assemble' task
tasks = ["assemble"]
}
clean_build {
title = "Clean Build"
versions = ["3.1", "/Users/me/gradle"]
tasks = ["build"]
gradle-args = ["--parallel"]
system-properties {
"key" = "value"
}
cleanup-tasks = ["clean"]
run-using = tooling-api // value can be "cli" or "tooling-api"
daemon = warm // value can be "warm", "cold", or "none"
measured-build-ops = ["org.gradle.api.internal.tasks.SnapshotTaskInputsBuildOperationType"] // see --measure-build-op
buck {
targets = ["//thing/res_debug"]
type = "android_binary" // can be a Buck build rule type or "all"
}
warm-ups = 10
}
ideaModel {
title = "IDEA model"
# Fetch the IDEA tooling model
tooling-api {
model = "org.gradle.tooling.model.idea.IdeaProject"
}
# Can also run tasks
# tasks = ["assemble"]
}
toolingAction {
title = "IDEA model"
# Fetch the IDEA tooling model
tooling-api {
action = "org.gradle.profiler.toolingapi.FetchProjectPublications"
}
# Can also run tasks
# tasks = ["assemble"]
}
androidStudioSync {
title = "Android Studio Sync"
# Measure an Android studio sync
# Note: Android Studio Bumblebee (2021.1.1) or newer is required
# Note2: If you are testing with Android Studio Giraffe (2022.3) or later
# you need to have local.properties file in your project with sdk.dir set
android-studio-sync {
# Override default Android Studio jvm args
# studio-jvm-args = ["-Xms256m", "-Xmx4096m"]
# Pass an IDEA properties to Android Studio. This can be used to set a registry values as well
# idea-properties = ["gradle.tooling.models.parallel.fetch=true"]
}
}
Values are optional and default to the values provided on the command-line or defined in the build.
A scenario can define changes that should be applied to the source before each build. You can use this to benchmark or profile an incremental build. The following mutations are available:
apply-build-script-change-to: Add a statement to a Groovy or Kotlin DSL build script, init script or settings script. Each iteration adds a new statement and removes the statement added by the previous iteration.apply-project-dependency-change-to: Add project dependencies to a Groovy or a Kotlin DSL build script. Each iteration adds a new combination of projects as dependencies and removes the projects added by the previous iteration.apply-abi-change-to: Add a public method to a Java or Kotlin source class. Each iteration adds a new method and removes the method added by the previous iteration.apply-non-abi-change-to: Change the body of a public method in a Java or Kotlin source class.apply-h-change-to: Add a function to a C/C++ header file. Each iteration adds a new function declaration and removes the function added by the previous iteration.apply-cpp-change-to: Add a function to a C/C++ source file. Each iteration adds a new function and removes the function added by the previous iteration.apply-property-resource-change-to: Add an entry to a properties file. Each iteration adds a new entry and removes the entry added by the previous iteration.apply-android-resource-change-to: Add a string resource to an Android resource file. Each iteration adds a new resource and removes the resource added by the previous iteration.apply-android-resource-value-change-to: Change a string resource in an Android resource file.apply-android-manifest-change-to: Add a permission to an Android manifest file.apply-android-layout-change-to: Add a hidden view with id to an Android layout file. Supports traditional layouts as well as Databinding layouts with a ViewGroup as the root element.apply-kotlin-composable-change-to: Add a @Composable function to a Kotlin source file.clear-build-cache-before: Deletes the contents of the build cache before the scenario is executed (SCENARIO), before cleanup (CLEANUP) or before the build is executed (BUILD).clear-gradle-user-home-before: Deletes the contents of the Gradle user home directory before the scenario is executed (SCENARIO), before cleanup (CLEANUP) or before the build is executed (BUILD).wrapper cache in the Gradle user home, since the downloaded wrapper in that location is used to run Gradle.none daemon option to use with CLEANUP or BUILD.clear-configuration-cache-state-before: Deletes the contents of the .gradle/configuration-cache-state directory before the scenario is executed (SCENARIO), before cleanup (CLEANUP) or before the build is executed (BUILD).clear-project-cache-before: Deletes the contents of the .gradle and buildSrc/.gradle project cache directories before the scenario is executed (SCENARIO), before cleanup (CLEANUP) or before the build is executed (BUILD).clear-transform-cache-before: Deletes the contents of the transform cache before the scenario is executed (SCENARIO), before cleanup (CLEANUP) or before the build is executed (BUILD).clear-jars-cache-before: Deletes the contents of the instrumented jars cache before the scenario is executed (SCENARIO), before cleanup (CLEANUP) or before the build is executed (BUILD).clear-android-studio-cache-before: Invalidates the Android Studio caches before the scenario is executed (SCENARIO) or before the build is executed (BUILD). Due to Android Studio client specifics before cleanup (CLEANUP) is not supported. Note: cleaning the Android Studio caches is run only when Android Studio sync (android-studio-sync) is used.git-checkout: Checks out a specific commit for the build step, and a different one for the cleanup step.git-revert: Reverts a given set of commits before the build and resets it afterward.iterations: Number of builds to actually measurejvm-args: Sets or overrides the jvm arguments set by org.gradle.jvmargs in gradle.properties.show-build-cache-size: Shows the number of files and their size in the build cache before scenario execution, and after each cleanup and build round…warm-ups: Number of warmups to perform before measurementThey can be added to a scenario file like this:
incremental_build {
tasks = ["assemble"]
apply-build-script-change-to = "build.gradle.kts"
apply-project-dependency-change-to {
files = ["build.gradle"]
# Default number of dependency-count is 3.
# Gradle Profiler will simulate changes to project dependencies by generate some additional projects and then add a combination of project dependencies to every non-generated subprojects before each iteration.
# The profiler will generate the minimal number of subprojects to allow for a unique combination of dependencies to be used for each iteration.
# Note: Number of generated projects is calculated as binomial coffiecient: "from `x` choose `dependency-count` = `iterations * files`", where number of generated projects is `x`.
dependency-count = 3
}
apply-abi-change-to = "src/main/java/MyThing.java"
apply-non-abi-change-to = ["src/main/java/MyThing.java", "src/main/java/MyOtherThing.java"]
apply-h-change-to = "src/main/headers/app.h"
apply-cpp-change-to = "src/main/cpp/app.cpp"
apply-property-resource-change-to = "src/main/resources/thing.properties"
apply-android-resource-change-to = "src/main/res/values/strings.xml"
apply-android-resource-value-change-to = "src/main/res/values/strings.xml"
apply-android-manifest-change-to = "src/main/AndroidManifest.xml"
clear-build-cache-before = SCENARIO
clear-transform-cache-before = BUILD
show-build-cache-size = true
git-checkout = {
cleanup = "efb43a1"
build = "master"
}
git-revert = ["efb43a1"]
jvm-args = ["-Xmx2500m", "-XX:MaxMetaspaceSize=512m"]
}
You can compare Gradle against Bazel, Buck, and Maven by specifying their equivalent invocations in the scenario file. Only benchmarking mode is supported.
> gradle-profiler --benchmark --maven clean_build
clean_build {
tasks = ["build"]
cleanup-tasks = ["clean"]
maven {
# If empty, it will be infered from MAVEN_HOME environment variable
home = "/path/to/maven/home"
targets = ["clean", "build"]
}
}
> gradle-profiler --benchmark --bazel build_some_target
build_some_target {
tasks = ["assemble"]
bazel {
# If empty, it will be infered from BAZEL_HOME environment variable
home = "/path/to/bazel/home"
targets = ["build" "//some/target"]
}
}
> gradle-profiler --benchmark --buck build_binaries
build_binaries {
tasks = ["assemble"]
buck {
# If empty, it will be infered from BUCK_HOME environment variable
home = "/path/to/buck/home"
type = "android_binary" // can be a Buck build rule type or "all"
}
}
build_resources {
tasks = ["thing:processDebugResources"]
buck {
targets = ["//thing/res_debug"]
}
}