performance-tools-main

Parent project for the agnostic performance-tools core and it's junit templates performance-tools-junit.

License

License

Categories

Categories

MINA Net Networking ORM Data
GroupId

GroupId

com.fillumina
ArtifactId

ArtifactId

performance-tools-main
Last Version

Last Version

1.1
Release Date

Release Date

Type

Type

pom
Description

Description

performance-tools-main
Parent project for the agnostic performance-tools core and it's junit templates performance-tools-junit.
Project URL

Project URL

http://maven.apache.org
Source Code Management

Source Code Management

http://github.com/fillumina/PerformanceTools

Download performance-tools-main

How to add to project

<!-- https://jarcasting.com/artifacts/com.fillumina/performance-tools-main/ -->
<dependency>
    <groupId>com.fillumina</groupId>
    <artifactId>performance-tools-main</artifactId>
    <version>1.1</version>
    <type>pom</type>
</dependency>
// https://jarcasting.com/artifacts/com.fillumina/performance-tools-main/
implementation 'com.fillumina:performance-tools-main:1.1'
// https://jarcasting.com/artifacts/com.fillumina/performance-tools-main/
implementation ("com.fillumina:performance-tools-main:1.1")
'com.fillumina:performance-tools-main:pom:1.1'
<dependency org="com.fillumina" name="performance-tools-main" rev="1.1">
  <artifact name="performance-tools-main" type="pom" />
</dependency>
@Grapes(
@Grab(group='com.fillumina', module='performance-tools-main', version='1.1')
)
libraryDependencies += "com.fillumina" % "performance-tools-main" % "1.1"
[com.fillumina/performance-tools-main "1.1"]

Dependencies

There are no dependencies for this project. It is a standalone project that does not depend on any other jars.

Project Modules

  • performance-tools
  • performance-tools-junit
  • performance-tools-examples
  • performance-tools-testng

Performance-Tools

A framework to easily perform comparative benchmarks and assertions on Java code.

Knowing how fast a certain code is on a specific machine (with a specific JDK, SO etc etc) tells almost nothing. I want to know if a code or an algorithm is faster or slower in respect of an alternative. Did I improved my code or is it slower? Is this the better solution under some conditions? That’s what I ask myself and that’s what this API is for: to be able to compare different codes together and see which one performs better. And I can even put this check in a unit test to make sure my assumptions stay correct during development!

Index

Note to present version

Added performance-tools-testng templates.

Use with maven

This project can be used with maven by adding the following dependencies to your project configuration pom.xml.

This is the core project, can be used alone but it's easier to use throught templates (just specify the one corresponding to the test unit framework you are using):

    <dependency>
        <groupId>com.fillumina</groupId>
        <artifactId>performance-tools</artifactId>
        <version>1.1</version>
    </dependency>

This module contains the templates for JUnit:

    <dependency>
        <groupId>com.fillumina</groupId>
        <artifactId>performance-tools-junit</artifactId>
        <version>1.1</version>
    </dependency>

This module contains the templates for TestNG:

    <dependency>
        <groupId>com.fillumina</groupId>
        <artifactId>performance-tools-testng</artifactId>
        <version>1.1</version>
    </dependency>

Because the templates depend on the core project you only need to specify the right template.

History

  • version 1.0 released 28 July 2014: first version released to maven central
  • version 0.1 released 4 October 2013

Summary

To asses performances of a java code can be tough: first Java runs on a variety of platforms (from mobiles to mainframes) and it's difficult to tell how some code will perform in all of them (think about memory or FPU constrains); then the JVM itself is not unique (there are many producers) and may execute the byte-code in different ways. Other than that multitasking, virtualization and even the Java garbage collector may impact the performance results unpredictably.

Evaluating how long a (relatively small) code takes to execute is a kind of performance test called micro-benchmark as opposed to the classic benchmarks which usually concern full program executions. A micro-benchmark has the disadvantage of making it difficult to compare measures obtained from different systems.

Another approach is to take the measurements of two or more different codes and consider the relative speed between them. This technique has the following advantages over a micro-benchmark:

  • The framework overhead is eliminated completely;
  • Percentages are more reproducible between different systems;
  • It's far more informative to know how much a code is faster in respect of another known code;
  • It's more robust against environment disturbances (CPU-time fluctuations).

This is the approach chosen by this API (although nothing forbids to specify only one test and so having a micro-benchmark).

It has the following features:

  • It allows to test codes in a single and in multi threaded environment;
  • It allows to specify parametrized codes;
  • It allows to use a sequence as a second parameter of a parametrized code;
  • It allows to easily export the performances or use them in place;
  • It allows to assert conditions on the tests so to use them in unit tests;
  • The structure of the API is very open and interface centric so that it is highly customizable and expandable;
  • It can be used with two different paradigms: fluent interface and templates.

Bibliography

Compilation and installation

This is a multi-artifact maven project so you can build and install the whole project by issuing

mvn clean install

from the main directory. The actual API core is in the folder 'performance-tools' and its jar should be in 'performance-tools/target'. The JUnit templates jar is in the folder 'performance-tools-junit/target'.

Failing tests

Please note that performance tests cannot be assured to be stable under any possible condition! In particular if the tolerance is strict you may expect test failure quite frequently. To overcome this problem relax your tolerances and try to run the performance test apart from the main building task. In particular, avoid to use test multitasking ('maven' can be configured to execute more than one test in parallel).

Usage

The easiest way to use this library is by extending one of its templates. To use a template is easy because the abstract methods are there to remind what it is needed and they are pretty self explanatory on how to do it (autocompletition should work with any decent IDE). Here is an example of a very simple JUnit performance test using a template that increases the iteration number at each round until it matches the required performance stability. The main() method is added so it can be executed by itself producing some useful output.

public class DivisionByTwoPerformanceTest
        extends JUnitAutoProgressionPerformanceTemplate {

    // allows to run the test stand alone with some useful output
    public static void main(final String[] args) {
        new DivisionByTwoPerformanceTest().testWithIntermediateOutput();
    }

    @Override
    public void init(ProgressionConfigurator config) {
        // this is where the target stability is set
        config.setMaxStandardDeviation(2);
    }

    @Override
    public void addTests(TestsContainer tests) {
        final Random rnd = new Random(System.currentTimeMillis());

        tests.addTest("math", new RunnableSink() {

            @Override
            public Object sink() {
                return rnd.nextInt() / 2;
            }
        });

        tests.addTest("binary", new RunnableSink() {

            @Override
            public Object sink() {
                return rnd.nextInt() >> 1;
            }
        });
    }

    @Override
    public void addAssertions(PerformanceAssertion assertion) {
        assertion.setPercentageTolerance(2)
                .assertTest("binary").fasterThan("math");
    }
}

This is the result returned by calling the test's main():

Iterations: 1000	Samples: 10	Standard Deviation: 7.997750553879185
Iterations: 10000	Samples: 10	Standard Deviation: 22.840778409983503
Iterations: 10000	Samples: 10	Standard Deviation: 3.628900909423828
Iterations: 100000	Samples: 10	Standard Deviation: 10.10749740600586
Iterations: 100000	Samples: 10	Standard Deviation: 9.441978610152637
Iterations: 1000000	Samples: 10	Standard Deviation: 3.9505724269289453
Iterations: 10000000	Samples: 10	Standard Deviation: 0.3524486164596204

 (10,000,000 iterations)
math  	   0 :	     20.09 ns		    100.00 %
binary	   1 :	     19.35 ns		     96.34 %
           * :	     39.44 ns

Note that some iterations are executed twice because there weren’t any improvement in the stability so the API automatically implies that some disturbance had occurred and repeated the iteration.

There are other templates to manage parameters and sequence:

  • The parameters are useful to test a single code against different similar objects (like different implementation of a Map to see which one is faster).
  • The sequence allows to test a parametrized code against various values (like different Map sizes to see how size impacts performances).

Each test can be executed in a multi threaded environment in a very easy way (by just modifying its configuration).

There are text and CSV viewers built-in, but if you need something different (a Swing GUI or to save values on disk) you may implement an interface and call the 'executeTest()' method of the templates.

Versions

Version
1.1
1.0