search

gradle / gradle-profiler

A tool for gathering profiling and benchmarking information for Gradle builds
Apache License 2.0
1.39k stars 156 forks source link
gradle-bt gradle-bt-core-execution

readme

Gradle Profiler

A tool to automate the gathering of profiling and benchmarking information for Gradle builds.

Profiling information can be captured using several different tools:

Installing

SDKMAN!

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

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

Download binaries

Binaries are available and linked from the releases page.

Build from source

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 a build

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.

Regression detection

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 a build

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

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

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:

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

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:

YourKit

YourKit 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.

Java Flight Recorder

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.

Heap dump

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 Trace

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.

Command line options

The following command line options only apply when measuring Gradle builds:

Advanced profiling scenarios

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.

Profiling incremental builds

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:

They 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
    copy-file = {
        source = "../develocity.xml"
        target = ".mvn/develocity.xml"
    }
    delete-file = [{
        target = ".mvn/develocity.xml"
        schedule = CLEANUP
    }, {
        target = ".gradle"
        schedule = CLEANUP
    }]
    git-checkout = {
        cleanup = "efb43a1"
        build = "master"
    }
    git-revert = ["efb43a1"]
    jvm-args = ["-Xmx2500m", "-XX:MaxMetaspaceSize=512m"]
}

Comparing against other build tools

You can compare Gradle against Bazel, Buck, and Maven by specifying their equivalent invocations in the scenario file. Only benchmarking mode is supported.

Maven

> 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"]
    }
}

Bazel

> 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"]
    }
}

Buck

> 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"]
    }
}