5.0 KiB
Micro BenchMark Notice
All optimization must be based on data verification, and blind optimization is rejected. Based on this, we provide the MicroBench module.
The MicroBench module is based on the OpenJDK JMH component (HotSpot's recommended benchmark test program). When you start benchmarking, you don't need additional dependencies.
JMH, the Java MicroBenchmark Harness, is a tool suite dedicated to code microbenchmark testing. What is Micro Benchmark? Simply put, it is based on method-level benchmark testing, with an accuracy of microseconds. When you locate a hot method and want to further optimize the performance of the method, you can use JMH to quantitatively analyze the optimized results.
Several points to note in Java benchmark testing:
-
Prevent useless code from entering the test method.
-
Concurrent testing.
-
The test results are presented.
Typical application scenarios of JMH are:
-
1: Quantitatively analyze the optimization effect of a hotspot function
-
2: Want to quantitatively know how long a function needs to be executed, and the correlation between execution time and input variables
-
3: Compare multiple implementations of a function
DolphinScheduler-MicroBench provides AbstractBaseBenchmark, you can inherit from it, write your benchmark code, AbstractMicroBenchmark can guarantee to run in JUnit mode.
Customized operating parameters
The default AbstractMicrobenchmark configuration is
Warmup times 10 (warmupIterations)
Number of tests 10 (measureIterations)
Fork quantity 2 (forkCount)
You can specify these parameters at startup，-DmeasureIterations, -DperfReportDir (output benchmark test result file directory), -DwarmupIterations, -DforkCount
DolphinScheduler-MicroBench Introduction
It is generally not recommended to use fewer cycles when running tests. However, a smaller number of tests helps to verify the work during the benchmark test. After the verification is over, run a large number of benchmark tests.
@Warmup(iterations = 2, time = 1)
@Measurement(iterations = 4, time = 1)
@State(Scope.Benchmark)
public class EnumBenchMark extends AbstractBaseBenchmark {
}
This can run benchmarks at the method level or the class level. Command line parameters will override the parameters on the annotation.
@Benchmark // Method annotation, indicating that the method is an object that needs to be benchmarked.
@BenchmarkMode(Mode.AverageTime) // Optional benchmark test mode is obtained through enumeration
@OutputTimeUnit(TimeUnit.MICROSECONDS) // Output time unit
public void enumStaticMapTest() {
TestTypeEnum.newGetNameByType(testNum);
}
When your benchmark test is written, you can run it to view the specific test conditions: (The actual results depend on your system configuration)
First, it will warm up our code,
# Warmup Iteration 1: 0.007 us/op
# Warmup Iteration 2: 0.008 us/op
Iteration 1: 0.004 us/op
Iteration 2: 0.004 us/op
Iteration 3: 0.004 us/op
Iteration 4: 0.004 us/op
After warmup, we usually get the following results
Benchmark (testNum) Mode Cnt Score Error Units
EnumBenchMark.simpleTest 101 thrpt 8 428750972.826 ± 66511362.350 ops/s
EnumBenchMark.simpleTest 108 thrpt 8 299615240.337 ± 290089561.671 ops/s
EnumBenchMark.simpleTest 103 thrpt 8 288423221.721 ± 130542990.747 ops/s
EnumBenchMark.simpleTest 104 thrpt 8 236811792.152 ± 155355935.479 ops/s
EnumBenchMark.simpleTest 105 thrpt 8 472247775.246 ± 45769877.951 ops/s
EnumBenchMark.simpleTest 103 thrpt 8 455473025.252 ± 61212956.944 ops/s
EnumBenchMark.enumStaticMapTest 101 avgt 8 0.006 ± 0.003 us/op
EnumBenchMark.enumStaticMapTest 108 avgt 8 0.005 ± 0.002 us/op
EnumBenchMark.enumStaticMapTest 103 avgt 8 0.006 ± 0.005 us/op
EnumBenchMark.enumStaticMapTest 104 avgt 8 0.006 ± 0.004 us/op
EnumBenchMark.enumStaticMapTest 105 avgt 8 0.004 ± 0.001 us/op
EnumBenchMark.enumStaticMapTest 103 avgt 8 0.004 ± 0.001 us/op
EnumBenchMark.enumValuesTest 101 avgt 8 0.011 ± 0.004 us/op
EnumBenchMark.enumValuesTest 108 avgt 8 0.025 ± 0.016 us/op
EnumBenchMark.enumValuesTest 103 avgt 8 0.019 ± 0.010 us/op
EnumBenchMark.enumValuesTest 104 avgt 8 0.018 ± 0.018 us/op
EnumBenchMark.enumValuesTest 105 avgt 8 0.014 ± 0.012 us/op
EnumBenchMark.enumValuesTest 103 avgt 8 0.012 ± 0.009 us/op
OpenJDK officially gave a lot of sample codes, interested students can query and learn JMH by themselves:OpenJDK-JMH-Example