Code for running benchmarks that assess the performance of the random generators and samplers in Commons RNG.
Build the package using:
mvn package -Pexamples-jmh
The Java Microbenchmark Harness (JMH) is used to run the benchmarks. Options to control the benchmark can be detailed using the help flag:
java -jar target/examples-jmh.jar -h
For example controlling the length of each benchmark, the number of iterations and warm-up iterations, or the results format.
Executing the jar with no options will run all benchmarks. This will take a long time. The source code is organised into packages that reflect the module structure of Commons RNG. Browse the source for core, sampling and simple to see what benchmarks are available. Benchmarks can be targeted by using a regular expression naming the benchmark class and optionally the method.
By default JMH will output results to the console. To save to file use the -rff option and specify the format using the -rf option.
Run a benchmark of the UniformRandomProvider.nextDouble method for all random providers saving the results to file in JSON format:
java -jar target/examples-jmh.jar NextDoubleGenerationPerformance -rf json -rff out.json
Run a benchmark of the UniformRandomProvider.nextInt method for the named random providers:
java -jar target/examples-jmh.jar NextIntGen.*nextInt$ -p randomSourceName=MSWS,PCG_MCG_XSH_RS_32
Run a benchmark of the ContinuousSampler.sample method for the named sampler using all random providers:
java -jar target/examples-jmh.jar ContinuousSamplersPerformance.sample -p samplerType=ZigguratNormalizedGaussianSampler
Run a benchmark of DiscreteUniformSampler.sample or UniformRandomProvider.nextInt(int) for a specified number of samples for a given range. As the number of samples increases the use of a sampler is preferred over the use of the raw provider due to optimisations allowed during sampler construction.
java -jar target/examples-jmh.jar DiscreteUniformSamplerGenerationPerformance -p samples=1000,10000 -p upperBound=155
Compare the speed of generation of Gaussian samples using the named random provider. The relative speed of the two implementations of the ziggurat method can vary significantly across JDK version and target platform. The ZigguratSampler.NormalisedGaussian is more consistent in performance and currently the preferred Gaussian sampler.
java -jar target/examples-jmh.jar NextGaussianPerformance -p randomSourceName=XO_RO_SHI_RO_128_PP