Java并行流性能

Question

在尝试新的 Java 流时，我注意到与并行流的性能相关的一些奇怪的事情。我使用了一个简单的程序，它从文本文件中读取单词并计算长度 > 5 的单词(测试文件有 30000 个单词):

    String contents = new String(Files.readAllBytes(Paths.get("text.txt")));
    List<String> words = Arrays.asList(contents.split("[\\P{L}]+"));
    long startTime;
    for (int i = 0; i < 100; i++) {
        startTime = System.nanoTime();
        words.parallelStream().filter(w -> w.length() > 5).count();
        System.out.println("Time elapsed [PAR]: " + (System.nanoTime() - startTime));
        startTime = System.nanoTime();
        words.stream().filter(w -> w.length() > 5).count();
        System.out.println("Time elapsed [SEQ]: " + (System.nanoTime() - startTime));
        System.out.println("------------------");
    }

这会在我的机器上生成以下输出(我只提到第一个和最后 5 个循环迭代):

Time elapsed [PAR]: 114185196
Time elapsed [SEQ]: 3222664
------------------
Time elapsed [PAR]: 569611
Time elapsed [SEQ]: 797113
------------------
Time elapsed [PAR]: 678231
Time elapsed [SEQ]: 414807
------------------
Time elapsed [PAR]: 755633
Time elapsed [SEQ]: 679085
------------------
Time elapsed [PAR]: 755633
Time elapsed [SEQ]: 393425
------------------
...
Time elapsed [PAR]: 90232
Time elapsed [SEQ]: 163785
------------------
Time elapsed [PAR]: 80396
Time elapsed [SEQ]: 154805
------------------
Time elapsed [PAR]: 83817
Time elapsed [SEQ]: 154377
------------------
Time elapsed [PAR]: 81679
Time elapsed [SEQ]: 186449
------------------
Time elapsed [PAR]: 68849
Time elapsed [SEQ]: 154804
------------------

为什么第一个处理比其他处理慢 100 倍？为什么并行流在第一次迭代中比顺序流慢，但在最后一次迭代中却快两倍？为什么顺序流和并行流随着时间的推移变得更快？这与循环优化有关吗？

后来编辑:根据 Luigi 的建议，我使用 JUnitBenchmarks 实现了基准测试。 :

List<String> words = null;

@Before
public void setup() {
    try {
        String contents = new String(Files.readAllBytes(Paths.get("text.txt")));
        words = Arrays.asList(contents.split("[\\P{L}]+"));
    } catch (IOException e) {
        e.printStackTrace();
    }
}

@BenchmarkOptions(benchmarkRounds = 100)
@Test
public void parallelTest() {
    words.parallelStream().filter(w -> w.length() > 5).count();
}

@BenchmarkOptions(benchmarkRounds = 100)
@Test
public void sequentialTest() {
    words.stream().filter(w -> w.length() > 5).count();
}

我还将测试文件的字数增加到 300000。新结果是:

Benchmark.sequentialTest: [measured 100 out of 105 rounds, threads: 1 (sequential)]

round: 0.08 [+- 0.04], round.block: 0.00 [+- 0.00], round.gc: 0.00 [+- 0.00], GC.calls: 62, GC.time: 1.53, time.total: 8.65, time.warmup: 0.81, time.bench: 7.85

Benchmark.parallelTest: [measured 100 out of 105 rounds, threads: 1 (sequential)]

round: 0.06 [+- 0.02], round.block: 0.00 [+- 0.00], round.gc: 0.00 [+- 0.00], GC.calls: 32, GC.time: 0.79, time.total: 6.82, time.warmup: 0.39, time.bench: 6.43

看来最初的结果是由错误的微基准配置引起的......

Answer 1

Hotspot JVM开始以解释模式执行程序，并在经过一些分析后将常用的部分编译为 native 代码。因此，循环的初始迭代通常很慢。