微基准测试
Netty 有一个名为“netty-microbench”的模块,它执行一系列微基准测试。它建立在 OpenJDK JMH 之上,这是 HotSpot 首选的微基准测试解决方案。它包含“内置电池”,因此您无需额外的依赖项即可开始使用。
您可以通过 maven 从命令行或直接在 IDE 中运行基准测试。要使用默认设置运行所有测试,请使用 mvn -DskipTests=false test
。您需要明确设置 skipTests=false
,因为我们不希望在常规测试运行期间将(可能耗时的)微基准测试作为单元测试执行。
如果一切顺利,您将看到 JMH 对 fork 数量执行预热和基准测试迭代,并向您展示一个简洁的摘要。以下是典型基准测试运行的样子(您将在输出中看到很多这样的运行)
# Fork: 2 of 2
# Warmup: 10 iterations, 1 s each
# Measurement: 10 iterations, 1 s each
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Throughput, ops/time
# Running: io.netty.microbench.buffer.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_1_0
# Warmup Iteration 1: 8454.103 ops/ms
# Warmup Iteration 2: 11551.524 ops/ms
# Warmup Iteration 3: 11677.575 ops/ms
# Warmup Iteration 4: 11404.954 ops/ms
# Warmup Iteration 5: 11553.299 ops/ms
# Warmup Iteration 6: 11514.766 ops/ms
# Warmup Iteration 7: 11661.768 ops/ms
# Warmup Iteration 8: 11667.577 ops/ms
# Warmup Iteration 9: 11551.240 ops/ms
# Warmup Iteration 10: 11692.991 ops/ms
Iteration 1: 11633.877 ops/ms
Iteration 2: 11740.063 ops/ms
Iteration 3: 11751.798 ops/ms
Iteration 4: 11260.071 ops/ms
Iteration 5: 11461.010 ops/ms
Iteration 6: 11642.912 ops/ms
Iteration 7: 11808.595 ops/ms
Iteration 8: 11683.780 ops/ms
Iteration 9: 11750.292 ops/ms
Iteration 10: 11769.986 ops/ms
Result : 11650.238 ±(99.9%) 229.698 ops/ms
Statistics: (min, avg, max) = (11260.071, 11650.238, 11808.595), stdev = 169.080
Confidence interval (99.9%): [11420.540, 11879.937]
最后,测试输出将类似于此(具体取决于您的系统设置和配置)
Benchmark Mode Samples Mean Mean error Units
i.n.m.b.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_1_0 thrpt 20 11658.812 120.728 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_2_256 thrpt 20 10308.626 147.528 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_3_1024 thrpt 20 8855.815 55.933 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_4_4096 thrpt 20 5545.538 1279.721 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_5_16384 thrpt 20 6741.581 75.975 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledDirectAllocAndFree_6_65536 thrpt 20 7252.869 70.609 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledHeapAllocAndFree_1_0 thrpt 20 9750.225 73.900 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledHeapAllocAndFree_2_256 thrpt 20 9936.639 657.818 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledHeapAllocAndFree_3_1024 thrpt 20 8903.130 197.533 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledHeapAllocAndFree_4_4096 thrpt 20 6664.157 74.163 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledHeapAllocAndFree_5_16384 thrpt 20 6374.924 337.869 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.pooledHeapAllocAndFree_6_65536 thrpt 20 6386.337 44.960 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledDirectAllocAndFree_1_0 thrpt 20 2137.241 30.792 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledDirectAllocAndFree_2_256 thrpt 20 1873.727 41.843 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledDirectAllocAndFree_3_1024 thrpt 20 1902.025 34.473 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledDirectAllocAndFree_4_4096 thrpt 20 1534.347 20.509 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledDirectAllocAndFree_5_16384 thrpt 20 838.804 12.575 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledDirectAllocAndFree_6_65536 thrpt 20 276.976 3.021 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledHeapAllocAndFree_1_0 thrpt 20 35820.568 259.187 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledHeapAllocAndFree_2_256 thrpt 20 19660.951 295.012 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledHeapAllocAndFree_3_1024 thrpt 20 6264.614 77.704 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledHeapAllocAndFree_4_4096 thrpt 20 2921.598 95.492 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledHeapAllocAndFree_5_16384 thrpt 20 991.631 49.220 ops/ms
i.n.m.b.ByteBufAllocatorBenchmark.unpooledHeapAllocAndFree_6_65536 thrpt 20 261.718 11.108 ops/ms
Tests run: 1, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 993.382 sec - in io.netty.microbench.buffer.ByteBufAllocatorBenchmark
您还可以直接从您的 IDE 运行基准测试。如果您已经导入了 netty 父项目,请打开 microbench
子项目并导航到 src/main/java/io/netty/microbench
命名空间。在 buffer
命名空间中,您可以像运行任何其他基于 JUnit 的测试一样运行 ByteBufAllocatorBenchmark
。主要区别在于(截至目前),您只能一次运行完整基准测试,而不能单独运行每个子基准测试。您应该在控制台中看到与直接通过 mvn
运行时看到相同的输出。
编写基准测试本身并不难,但写好它却很难。这并不是因为 microbench 项目难以使用,而是因为您需要在编写基准测试时避免常见的陷阱。值得庆幸的是,JMH 套件提供了有用的注释和功能来减轻其中大部分陷阱。要开始编写基准测试,您需要让您的基准测试扩展 AbstractMicrobenchmark
,这将确保测试通过 JUnit 运行并配置一些默认值
public class MyBenchmark extends AbstractMicrobenchmark {
}
下一步是创建一个使用 @GenerateMicroBenchmark
注释的方法(并为其提供一个描述性名称)
@GenerateMicroBenchmark
public void measureSomethingHere() {
}
现在最好的办法是查看 此处,以获取有关如何编写适当的 JMH 测试的示例和灵感。此外,请查看 JMH 主要作者之一的 演讲。
默认设置(如 AbstractMicrobenchmark
中所述)为
- 预热迭代:10
- 测量迭代:10
- 分叉数:2
这些设置可以通过运行时的系统属性(warmupIterations
、measureIterations
和 forks
)进行自定义
mvn -DskipTests=false -DwarmupIterations=2 -DmeasureIterations=3 -Dforks=1 test
请注意,通常不建议使用这么少的迭代,但有时这有助于查看基准测试是否有效,然后在稍后运行全面的基准测试。
请注意,您还可以通过注释逐个测试地自定义这些默认设置
@Warmup(iterations = 20)
@Fork(1)
public class MyBenchmark extends AbstractMicrobenchmark {
}
这可以在逐个类和逐个方法(基准测试)的基础上进行。请注意,命令行参数始终会覆盖注释默认值。