Microbenchmarks for the Array64 library for huge arrays on the JVM.
Table of contents generated with markdown-toc
These benchmarks were developed using OpenJDK's Java Microbenchmarking Harness.
These benchmarks compare the performance of Array64 against FastUtil's BigArrays module, a lower-level tool for manipulating 2D arrays in Java used in scientific and high-performance computing applications (and internally by Array64 for some operations).
The Maven pom.xml file for this project is a good template for anyone looking to write JMH microbenchmarks in Kotlin.
| Benchmark | Mode | Cnt | Score | Error | Units |
|---|---|---|---|---|---|
| FastByteArray64Benchmarks.sequentialAccessViaForEach | avgt | 25 | 1280.953 | ± 11.985 | ms/op |
| FastByteArray64Benchmarks.sequentialAccessViaIterator | avgt | 25 | 1352.156 | ± 13.178 | ms/op |
| FastUtilByteArrayBenchmarks.sequentialAccess | avgt | 25 | 1003.542 | ± 11.602 | ms/op |
| FastObjectArray64Benchmarks.sequentialAccessViaForEach | avgt | 25 | 1287.035 | ± 9.112 | ms/op |
| FastObjectArray64Benchmarks.sequentialAccessViaIterator | avgt | 25 | 1346.005 | ± 12.665 | ms/op |
| FastUtilObjectArrayBenchmarks.sequentialAccess | avgt | 25 | 1009.380 | ± 11.395 | ms/op |
| FastByteArray64Benchmarks.randomAccess | thrpt | 25 | 7515.261 | ± 33.293 | ops/ms |
| FastUtilByteArrayBenchmarks.randomAccess | thrpt | 25 | 7470.478 | ± 44.760 | ops/ms |
| FastObjectArray64Benchmarks.randomAccess | thrpt | 25 | 6715.841 | ± 114.816 | ops/ms |
| FastUtilObjectArrayBenchmarks.randomAccess | thrpt | 25 | 6673.575 | ± 53.757 | ops/ms |
These results are current as of Array64 v1.0.1.
The raw console output from running the benchmarks is available in raw/benchmarks-output.txt.
For the random access benchmarks, a single random element access counts as an operation. In this metric the performance difference between Array64 and FastUtil is negligible.
For the sequential access benchmarks, a single iteration over an array of 268,435,456 (228) elements counts as an operation. In this metric, Array64 performs around 25% worse than FastUtil. However, either library is sufficiently fast for most use-cases; it takes ~1.3s to iterate over more than a quarter of a billion elements, compared to ~1.0s.
These results were obtained on a machine with a Ryzen 7 3700x CPU and 64GB of RAM.
- Ensure Java and Maven are installed, and that you have at least 16GB RAM available for the application
- Clone this repository
- From the root directory of the project (with the pom.xml), run "mvn install package"
- Execute the JMH-generated executable jar file with the command "java -jar ./target/benchmarks.jar"
- Do something else for a few hours