Tech and Media Labs
This site uses cookies to improve the user experience.




JMH - Java Microbenchmark Harness

Jakob Jenkov
Last update: 2015-09-16

JMH is short for Java Microbenchmark Harness. JMH is a toolkit that helps you implement Java microbenchmarks correctly. JMH is developed by the same people who implement the Java virtual machine, so these guys know what they are doing. This JMH tutorial will teach you how to implement and run Java microbenchmarks with JMH.

Why Are Java Microbenchmarks Hard?

Writing benchmarks that correctly measure the performance of a small part of a larger application is hard. There are many optimizations that the JVM or underlying hardware may apply to your component when the benchmark executes that component in isolation. These optimizations may not be possible to apply when the component is running as part of a larger application. Badly implemented microbenchmarks may thus make you believe that your component's performance is better than it will be in reality.

Writing a correct Java microbenchmark typically entails preventung the optimizations the JVM and hardware may apply during microbenchmark execution which could not have been applied in a real production system. That is what JMH - the Java Microbenchmark Harness - is helping you do.

Getting Started With JMH

The easiest way to get started with JMH is to generate a new JMH project using the JMH Maven archetype. The JMH Maven archetype will generate a new Java project with a single, example benchmark Java class, and a Maven pom.xml file. The Maven pom.xml file contains the correct dependencies to compile and build your JMH microbenchmark suite.

Here is the Maven command line needed to generate a JMH project template:

 mvn archetype:generate
          -DinteractiveMode=false
          -DarchetypeGroupId=org.openjdk.jmh
          -DarchetypeArtifactId=jmh-java-benchmark-archetype
          -DgroupId=com.jenkov
          -DartifactId=first-benchmark
          -Dversion=1.0

This is one long command. There should be no line breaks in the command. I just added them to make the command easier to read.

This command line will create a new directory named first-benchmark (the artifactId specified in the Maven command). Inside this directory will be generated a new Maven source directory structure (src/main/java). Inside the java source root directory will be generated a single Java packaged named com.jenkov (actually a package named com with a subpackage named jenkov). Inside the com.jenkov package will be a JMH benchmark Java class named MyBenchmark.

Your First JMH Benchmark

It is time to write your first JMH benchmark class, or at least see how it is done.

The generated MyBenchmark class is a JMH class template which you can use to implement your JMH benchmarks. You can either implement your benchmarks directly in the generated MyBenchmark class, or create a new class in the same Java package. To make it easy for you to write your first JMH benchmark I will just use the generated class in this example.

Here is first how the generated MyBenchmark class looks:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;

public class MyBenchmark {

    @Benchmark
    public void testMethod() {
        // This is a demo/sample template for building your JMH benchmarks. Edit as needed.
        // Put your benchmark code here.
    }

}

You can put the code you want to measure inside the testMethod() method body. Here is an example:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;

public class MyBenchmark {

    @Benchmark
    public void testMethod() {
        // This is a demo/sample template for building your JMH benchmarks. Edit as needed.
        // Put your benchmark code here.

        int a = 1;
        int b = 2;
        int sum = a + b;
    }

}

Note: This particular example is a bad benchmark implementation, as the JVM can see that sum is never used, and may thus eliminate the sum calculation. Well in fact the whole method body could be removed by JVM dead code elimination. For now, just imagine that the testMethod() body actually contained a good benchmark implementation. I will get back how to implement better benchmarks with JMH later in this tutorial.

Building Your JMH Benchmark

You can now compile and build a benchmark JAR file from your JMH benchmark project using this Maven command:

mvn clean install

This Maven command must be executed from inside the generated benchmark project directory (in this example the first-benchmark directory).

When this command is executed a JAR file will be created inside the first-benchmark/target directory. The JAR file will be named benchmarks.jar

The benchmarks.jar File

When you build your JMH benchmarks, Maven will always generate a JAR file named benchmarks.jar in the target directory (Maven's standard output directory).

The benchmarks.jar file contains everything needed to run your benchmarks. It contains your compiled benchmark classes as well as all JMH classes needed to run the benchmark.

If your benchmarks has any external dependencies (JAR files from other projects needed to run your benchmarks), declare these dependencies inside the Maven pom.xml, and they will be included in the benchmarks.jar too.

Since benchmarks.jar is fully self contained, you can copy that JAR file to another computer to run your JMH benchmarks on that computer.

Running Your JMH Benchmarks

Once you have built your JMH benchmark code you can run the benchmark using this Java command:

java -jar target/benchmarks.jar

This will start JMH on your benchmark classes. JMH will scan through your code and find all benchmarks and run them. JMH will print out the results to the command line.

Running the benchmarks will take some time. JMH makes several warm ups, iterations etc. to make sure the results are not completely random. The more runs you have, the better average performance and high / low performance information you get.

You should let the computer alone while it runs the benchmarks, and you should close all other applications (if possible). If your computer is running other applications, these applications may take time from the CPU and give incorrect (lower) performance numbers.

JMH Benchmark Modes

JMH can run your benchmarks in different modes. The benchmark mode tells JMH what you want to measure. JMH offer these benchmark modes:

Throughput Measures the number of operations per second, meaning the number of times per second your benchmark method could be executed.
Average Time Measures the average time it takes for the benchmark method to execute (a single execution).
Sample Time Measures how long time it takes for the benchmark method to execute, including max, min time etc.
Single Shot Time Measures how long time a single benchmark method execution takes to run. This is good to test how it performs under a cold start (no JVM warm up).
All Measures all of the above.

The default benchmark mode is Throughput.

You specify what benchmark mode your benchmark should use with the JMH annotation BenchmarkMode. You put the BenchmarkMode annotation on top of your benchmark method. Here is a JMH BenchmarkMode example:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;
import org.openjdk.jmh.annotations.BenchmarkMode;
import org.openjdk.jmh.annotations.Mode;

public class MyBenchmark {

    @Benchmark @BenchmarkMode(Mode.Throughput)
    public void testMethod() {
        // This is a demo/sample template for building your JMH benchmarks. Edit as needed.
        // Put your benchmark code here.

        int a = 1;
        int b = 2;
        int sum = a + b;
    }

}

Notice the @BenchmarkMode(Mode.Throughput) annotation above the testMethod() method. That annotation specifies the benchmark mode. The Mode class contains constants for each possible benchmark mode.

Benchmark Time Units

JMH enables you to specify what time units you want the benchmark results printed in. The time unit will be used for all benchmark modes your benchmark is executed in.

You specify the benchmark time unit using the JMH annotation @OutputTimeUnit. The @OutputTimeUnit annotation takes a java.util.concurrent.TimeUnit as parameter to specify the actual time unit to use. Here is a JMH @OutputTimeUnit annotation example:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;
import org.openjdk.jmh.annotations.BenchmarkMode;
import org.openjdk.jmh.annotations.Mode;
import org.openjdk.jmh.annotations.OutputTimeUnit;

import java.util.concurrent.TimeUnit;

public class MyBenchmark {

    @Benchmark @BenchmarkMode(Mode.Throughput) @OutputTimeUnit(TimeUnit.MINUTES)
    public void testMethod() {
        // This is a demo/sample template for building your JMH benchmarks. Edit as needed.
        // Put your benchmark code here.

        int a = 1;
        int b = 2;
        int sum = a + b;
    }

}

In this example the time unit specified is minutes. This means that you want the output shown using the time unit minutes (e.g. operations per minute).

The TimeUnit class contains the following time unit constants:

  • NANOSECONDS
  • MICROSECONDS
  • MILLISECONDS
  • SECONDS
  • MINUTES
  • HOURS
  • DAYS

Benchmark State

Sometimes you way want to initialize some variables that your benchmark code needs, but which you do not want to be part of the code your benchmark measures. Such variables are called "state" variables. State variables are declared in special state classes, and an instance of that state class can then be provided as parameter to the benchmark method. This may sound a bit complicated, so here is a JMH benchmark state example:

package com.jenkov;

import org.openjdk.jmh.annotations.*;

import java.util.concurrent.TimeUnit;


public class MyBenchmark {

    @State(Scope.Thread)
    public static class MyState {
        public int a = 1;
        public int b = 2;
        public int sum ;
    }


    @Benchmark @BenchmarkMode(Mode.Throughput) @OutputTimeUnit(TimeUnit.MINUTES)
    public void testMethod(MyState state) {
        state.sum = state.a + state.b;
    }

}

In this example I have added a nested static class named MyState. The MyState class is annotated with the JMH @State annotation. This signals to JMH that this is a state class. Notice that the testMethod() benchmark method now takes an instance of MyState as parameter.

Notice also that the testMethod() body has now been changed to use the MyState object when performing its sum calculation.

State Scope

A state object can be reused across multiple calls to your benchmark method. JMH provides different "scopes" that the state object can be reused in. There state scope is specified in the parameter of the @State annotation. In the example above the scope chosen was Scope.Thread

The Scope class contains the following scope constants:

Thread Each thread running the benchmark will create its own instance of the state object.
Group Each thread group running the benchmark will create its own instance of the state object.
Benchmark All threads running the benchmark share the same state object.

State Class Requirements

A JMH state class must obey the following rules:

  • The class must be declared public
  • If the class is a nested class, it must be declared static (e.g. public static class ...)
  • The class must have a public no-arg constructor (no parameters to the constructor).

When these rules are obeyed you can annotate the class with the @State annotation to make JMH recognize it as a state class.

State Object @Setup and @TearDown

You can annotate methods in your state class with the @Setup and @TearDown annotations. The @Setup annotation tell JMH that this method should be called to setup the state object before it is passed to the benchmark method. The @TearDown annotation tells JMH that this method should be called to clean up ("tear down") the state object after the benchmark has been executed.

The setup and tear down execution time is not included in the benchmark runtime measurements.

Here is a JMH state object example that shows the use of the @Setup and @TearDown annotations:

package com.jenkov;

import org.openjdk.jmh.annotations.*;

import java.util.concurrent.TimeUnit;


public class MyBenchmark {

    @State(Scope.Thread)
    public static class MyState {

        @Setup(Level.Trial)
        public void doSetup() {
            sum = 0;
            System.out.println("Do Setup");
        }

        @TearDown(Level.Trial)
        public void doTearDown() {
            System.out.println("Do TearDown");
        }

        public int a = 1;
        public int b = 2;
        public int sum ;
    }

    @Benchmark @BenchmarkMode(Mode.Throughput) @OutputTimeUnit(TimeUnit.MINUTES)
    public void testMethod(MyState state) {
        state.sum = state.a + state.b;
    }
}

Notice the two new methods in the MyState class named doSetup() and doTearDown(). These methods are annotated with the @Setup and @TearDown annotations. This example only show two methods, but you could have more methods annotated with @Setup and @TearDown.

Notice also that the annotations take a parameter. There are three different values this parameter can take. The value you set instruct JMH about when the method should be called. The possible values are:

Level.Trial The method is called once for each time for each full run of the benchmark. A full run means a full "fork" including all warmup and benchmark iterations.
Level.Iteration The method is called once for each iteration of the benchmark.
Level.Invocation The method is called once for each call to the benchmark method.

If you have any doubts about when a setup or tear down method is called, try inserting a System.out.println() statement in the method. Then you will see. Then you can change the @Setup and @TearDown() parameter values until your setup and tear down methods are called at the right time.

Writing Good Benchmarks

Now that you have seen how to use JMH to write benchmarks, it is time to discuss how to write good benchmarks. As mentioned in the beginning of this JMH tutorial there are a couple of pitfalls that you can easily fall into when implementing benchmarks. I will discuss some of these pitfalls in the following sections.

One common pitfall is that the JVM may apply optimizations to your components when executed inside the benchmark which could not have been applied if the component was executed inside your real application. Such optimizations will make your code look faster than it will be in reality. I will discuss some of these optimizations later.

Loop Optimizations

It is tempting to put your benchmark code inside a loop in your benchmark methods, in order to repeat it more times per call to the benchmark method (to reduce the overhead of the benchmark method call). However, the JVM is very good at optimizing loops, so you may end up with a different result than what you expected. In general you should avoid loops in your benchmark methods, unless they are part of the code you want to measure (and not around the code you want to measure).

Dead Code Elimination

One of the JVM optimizations to avoid when implementing performance benchmarks is dead code elimination. If the JVM detects that the result of some computation is never used, the JVM may consider this computation dead code and eliminate it. Look at this benchmark example:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;

public class MyBenchmark {

    @Benchmark
    public void testMethod() {
        int a = 1;
        int b = 2;
        int sum = a + b;
    }

}

The JVM can detect that the calculation of a + b which is assigned to sum is never used. Therefore the JVM can remove the calculation of a + b altogether. It is considered dead code. The JVM can then detect that the sum variable is never used, and that subsequently a and b are never used. They too can be eliminated.

In the end, there is no code left in the benchmark. The results from running this benchmark are thus highly misleading. The benchmarks do not actually measure the time of adding two variables and assigning the value to a third variable. The benchmarks measure nothing at all.

Avoiding Dead Code Elimination

To avoid dead code elimination you must make sure that the code you want to measure does not look like dead code to the JVM. There are two ways to do that.

  • Return the result of your code from the benchmark method.
  • Pass the calculated value into a "black hole" provided by JMH.

I will show you examples of both methods in the following sections.

Return Value From Benchmark Method

Returning a computed value from the JMH benchmark method could look like this:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;

public class MyBenchmark {

    @Benchmark
    public int testMethod() {
        int a = 1;
        int b = 2;
        int sum = a + b;

        return sum;
    }

}

Notice how the testMethod() method now returns the sum variable. This way the JVM cannot just eliminate the addition, because the return value might be used by the caller. JMH will take of tricking the JVM into believing that the return value is actually used.

If your benchmark method is calculating multiple values that might end up being eliminated as dead code, you can either combine the two values into a single, and return that value (e.g. an object with both values in).

Passing Value to a Black Hole

An alternative to returning a combined value is to pass the calculated values (or returned / generated objects or whatever the result of your benchmark is) into a JMH black hole. Here is how passing values into a black hole looks:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;
import org.openjdk.jmh.infra.Blackhole;

public class MyBenchmark {

    @Benchmark
   public void testMethod(Blackhole blackhole) {
        int a = 1;
        int b = 2;
        int sum = a + b;
        blackhole.consume(sum);
    }
}

Notice how the testMethod() benchmark method now takes a Blackhole object as parameter. This will be provided to the test method by JMH when called.

Notice also how the calculated sum in the sum variable is now passed to the consume() method of the Blackhole instance. This will fool the JVM into thinking that the sum variable is actually being used.

If your benchmark method produces multiple results you can pass each of these results to a black hole, meaning calling consume() on the Blackhole instance for each value.

Constant Folding

Constant folding is another common JVM optimization. A calculation which is based on constants will often result in the exact same result, regardless of how many times the calculation is performed. The JVM may detect that, and replace the calculation with the result of the calculation.

As an example, look at this benchmark:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;

public class MyBenchmark {

    @Benchmark
    public int testMethod() {
        int a = 1;
        int b = 2;
        int sum = a + b;

        return sum;
    }

}

The JVM may detect that the value of sum is based on the two constant values 1 and 2 in a and b. It may thus replace the above code with this:

package com.jenkov;

import org.openjdk.jmh.annotations.Benchmark;

public class MyBenchmark {

    @Benchmark
    public int testMethod() {
        int sum = 3;

        return sum;
    }

}

Or even just return 3; directly. The JVM could even continue and never call the testMethod() because it knows it always returns 3, and just inline the constant 3 wherever the testMethod() was to be called.

Avoiding Constant Folding

To avoid constant folding you must not hardcode constants into your benchmark methods. Instead, the input to your calculations should come from a state object. This makes it harder for the JVM to see that the calculations are based on constant values. Here is an example:

package com.jenkov;

import org.openjdk.jmh.annotations.*;

public class MyBenchmark {

    @State(Scope.Thread)
    public static class MyState {
        public int a = 1;
        public int b = 2;
    }


    @Benchmark 
    public int testMethod(MyState state) {
        int sum = state.a + state.b;
        return sum;
    }
}

Remember, if your benchmark method calculates multiple values you can pass them through a black hole instead of returning them, to also avoid the dead code elimination optimization. For instance:

    @Benchmark
    public void testMethod(MyState state, Blackhole blackhole) {
        int sum1 = state.a + state.b;
        int sum2 = state.a + state.a + state.b + state.b;

        blackhole.consume(sum1);
        blackhole.consume(sum2);
    }

Jakob Jenkov




Copyright  Jenkov Aps
Close TOC