Post

Java Microbenchmark Harness (JMH)

Creating the first JMH project. A quick hands-on lesson to learn about Java Microbenchmark Harness (JMH). The article helps you get started and configure JMH project.

Introduction

In my previous article I established that microbenchmarking is hard with jvm. It is not enough to surround the code in a loop with System.out.println() and gather the time measurements. While benchmarking, a developer should consider warm-up cycles, JIT compilations, jvm optimizations, avoiding usual pitfalls and even more.

Thankfully, OpenJDK has a great tool Java Microbenchmark Harness (JMH) that can help to generated benchmarking stats. In this article, I will examine how JMH can help to avoid the pitfalls that we have discussed earlier.

Getting Started with JMH

A quick way to start with JMH is to use the Maven archetype. The command below will generate a new Java project. The project will have com/gaurav/MyBenchmark.java class and pom.xml. The Maven pom.xml includes all the required dependencies to support JMH.

1
mvn archetype:generate -DarchetypeGroupId=org.openjdk.jmh -DarchetypeArtifactId=jmh-java-benchmark-archetype -DinteractiveMode=false -DgroupId=com.gaurav -DartifactId=benchmark -Dversion=1.0

Good Benchmarks with JMH

Below are few features of JMH that help write better microbenchmarks.

  • JMH, by default, makes several warm up cycles before collecting the stats. Thus, it makes sure that the results are not completely random and jvm has performed initial optimizations.
  • @benchmark runs iteration over the code, before collecting the average. The more runs it makes through the code, the better stats it will collect.
  • Use Blackhole class of JMH to avoid dead code elimination by jvm. If I pass the calculated results to blackhole.consume(), it would trick the jvm. jvm will never drop the code thinking that consume() method uses the result.

Writing First Benchmark

Maven has already provided me with a template in MyBenchmark class to fill in. I am going to utilise the same class.

1
2
3
4
5
6
7
8
9
10
package com.gaurav;
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.
    }
}

I would like to keep my first benchmark pretty simple. Let me start by iterating over all the elements of a list and sum them up using a conventional for loop. As discussed, I will use Blackhole to fool the compiler and return the result. Here, I am asking JMH to calculate the average time, using @BenchmarkMode, which it takes to run the testMethod().

1
2
3
4
5
6
7
8
9
10
11
@Benchmark
@BenchmarkMode(Mode.AverageTime)
public static double testMethod(Blackhole blackhole) {
    double sum = 0;
    for(int i=0; i<list.size(); i++) {
        sum += list.get(i);
    }

    blackhole.consume(sum);
    return sum;
}

Compiling the JMH Project

Compile and build the project like any other Maven project:

1
mvn clean install

The command will create a fully executable jar file under benchmark/target directory. Please note that Maven will always generate a jar file named benchmarks.jar, regardless of the project name.

The next step is to execute the jar.

1
java -jar target/benchmarks.jar

Executing above command produced below result for me. It means that test operation is taking approx. 0.053 seconds on the current hardware.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# Run progress: 80.00% complete, ETA 00:01:41
# Fork: 5 of 5
# Warmup Iteration   1: 0.052 s/op
# Warmup Iteration   2: 0.051 s/op
# Warmup Iteration   3: 0.053 s/op
# Warmup Iteration   4: 0.056 s/op
# Warmup Iteration   5: 0.055 s/op
Iteration   1: 0.054 s/op
Iteration   2: 0.053 s/op
Iteration   3: 0.053 s/op
Iteration   4: 0.054 s/op
Iteration   5: 0.059 s/op

Result "com.example.MyBenchmark.testMethod":
  0.053 ±(99.9%) 0.002 s/op [Average]
  (min, avg, max) = (0.052, 0.053, 0.061), stdev = 0.002
  CI (99.9%): [0.051, 0.055] (assumes normal distribution)

# Run complete. Total time: 00:08:27

Benchmark Modes

In the previous example, I used @BenchmarkMode(Mode.AverageTime). If you try to decompile JMH jar, you will find enum Mode has below options:

Modes 
Throughput("thrpt", "Throughput, ops/time")It will calculate the number of times your method can be executed with in a second
AverageTime("avgt", "Average time, time/op")It will calculate the average time in seconds to execute the test method
SampleTime("sample", "Sampling time")It randomly samples the time spent in test method calls
SingleShotTime("ss", "Single shot invocation time")It works on single invocation of the method and is useful in calculating cold performance
All("all", "All benchmark modes")Calculates all of the above

The default Mode is Throughput.

Time measurement

It is evident from the console output above that calculations are in seconds. But, JMH allows to configure the time units using @OutputTimeUnit annotation. The @OutputTimeUnit accepts java.util.concurrent.TimeUnit, as shown below:

1
@OutputTimeUnit(TimeUnit.SECONDS)

The TimeUnit enum has following values:

NANOSECONDS
MICROSECONDS
MILLISECONDS
SECONDS
MINUTES
HOURS
DAYS

The default TimeUnit is SECONDS

Configure Fork, Warmup and Iterations

The benchmark is currently executing 5 times, with 5 warmup iterations and 5 measurement iterations. JMH even allows to configure these values using @Fork, @Warmup and @Measurement annotations. The code snippet below would execute the test method twice, with a couple of warmup iterations and 3 measurement iterations.

1
2
3
@Fork(value = 2)
@Warmup(iterations = 2)
@Measurement(iterations = 3)

@Warmup and @Measurement annotations also accepts parameters:

  • batchSize - configures the number of test method calls to be performed per operation
  • time - time spent for each iteration

Practice

You can play around to compare execution times of different for loops i.e. a conventional for loop, a forEach loop and a stream iterator. Something like:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
private static final List<Integer> list = IntStream.rangeClosed(1, Integer.MAX_VALUE/100)
            .boxed().collect(Collectors.toList());

    @Benchmark
    @BenchmarkMode(Mode.AverageTime)
    public static double conventionalLoop(Blackhole blackhole) {
        double sum = 0;
        for(int i=0; i<list.size(); i++) {
            sum += list.get(i);
        }
        
        blackhole.consume(sum);
        return sum;
    }

    @Benchmark
    @BenchmarkMode(Mode.AverageTime)
    public static double enhancedForLoop(Blackhole blackhole) throws InterruptedException {
        double sum = 0;
        for (int integer : list) {
            sum += integer;
        }

        blackhole.consume(sum);
        return sum;
    }

    @Benchmark
    @BenchmarkMode(Mode.AverageTime)
    public static double streamMap(Blackhole blackhole) {
        double sum = list.stream().mapToDouble(Integer::doubleValue).sum();
        blackhole.consume(sum);
        return sum;
    }

Conclusion

In this post, we have gone through a hands-on example of creating a JMH project. We have seen how can we configure our JMH project to suit our needs. You can refer to JMH Github Samples for more in depth examples.

We have seen that JMH is a jvm tool. In the next article we will try to explore if it can help us with other jvm based languages.

Reference
JMH Github
JMH Github Samples
JMH Javadox - Mode
JMH Javadox - OutputTimeUnit
JMH Javadox - Fork

This post is licensed under CC BY 4.0 by the author.

Comments powered by Disqus.