💪🏿 🤔 👰🏾 JBreakを使用したパフォーマンスタスクの分析（パート1） 👩🏻‍🤝‍👨🏿 🚠 🕚

最初の部分は、4つの最もホリバー問題の分析です。

void forEach(List<Integer> values, PrintStream ps) { values.forEach(ps::println); } void forEach(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } void forEach(List<Integer> values, PrintStream ps) { values.parallelStream().forEach(ps::println); }

カットの下で、タスクの状態、それらの外観の履歴、最初のタスクの分析、および会議参加者間の正しいソリューションの統計。

シリーズの他の出版物：パート2 、パート3 、パート4 。

まえがき

カンファレンスの10日前の同僚が、抽選のタスクを準備する提案を思い付きました。そこでは、Joshua Blochによる複数のEffective Java本（最後の第3版）が最優秀賞になります。そのような賞品の下では、適切なタスクが必要です。そして、呼び出しは受け入れられます。

まず、取得するものを決定します。

タスクはコード上にある必要があります（私は論理的なタスクを作成するマスターではありませんが、準備ができているタスクは取りたくありませんでした）
タスクに明らかな間違った答えがなければなりません
タスクはすべて単純またはすべて複雑である必要はありません
タスクをエンコードする場合、100％の答えは得られませんが、適切なソリューションを見つけるのに役立ちます
理想的には、会議の後に真実の底に到達したいという願望があるように、タスクは「フック」されるべきです。

そこで、次の形式でパフォーマンスタスクを準備することにしました。

同様のことを行うアルゴリズムのいくつかのバリアント
速度で並べ替える必要があります
アルゴリズムがこのように順序付けられている理由を説明します。

タスク

問題の状態が印刷でどのように見えたか

すべてのタスクに関する解説：

作業の速度でメソッドを並べ替え、適切なセルにメソッドの数を入力します。
メソッドの結果が少し異なる場合は、1つのセルに入力してください。
各タスクについて、一部のメソッドが他のメソッドよりも速い理由の説明を書きます。

タスク番号1

  void forEach(List<Integer> values, PrintStream ps) { values.forEach(ps::println); } void forEach(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } void forEach(List<Integer> values, PrintStream ps) { values.parallelStream().forEach(ps::println); }

タスク番号2

  String format(String user, String grade, String company, String message) { return String.format( ", %s,   %s .    %s    %s", user, grade, company, message); } String format(String user, String grade, String company, String message) { return ", " + user + ",   " + grade + " .    " + company + "    " + message; } String format(String user, String grade, String company, String message) { return new StringBuilder(", ") .append(user) .append(",   ") .append(grade) .append(" .    ") .append(company) .append("    ") .append(message) .toString(); }

タスク番号3

  public static double compute( double x1, double y1, double z1, double x2, double y2, double z2) { double x = y1 * z2 - z1 * y2; double y = z1 * x2 - x1 * z2; double z = x1 * y2 - y1 * x2; return x * x + y * y + z * z; } public static double compute( double x1, double y1, double z1, double x2, double y2, double z2) { Vector v1 = new Vector(x1, y1, z1); Vector v2 = new Vector(x2, y2, z2); return v1.crossProduct(v2).squared(); } public final static class Vector { private final double x, y, z; public Vector(double x, double y, double z) { this.x = x; this.y = y; this.z = z; } public double squared() { return x * x + y * y + z * z; } public Vector crossProduct(Vector v) { return new Vector( y * vz - z * vy, z * vx - x * vz, x * vy - y * vx); } }

タスク番号4

  public double octaPow(double a) { return Math.pow(a, 8); } public double octaPow(double a) { return a * a * a * a * a * a * a * a; } public double octaPow(double a) { return Math.pow(Math.pow(Math.pow(a, 2), 2), 2); } public double octaPow(double a) { a *= a; a *= a; return a * a; }

コードはHotSpot 64ビットVM（JRE 1.8.0_161）で実行されることを前提としています。 JREの選択されたバージョンに固有のものはありませんが、それは最後のバージョンの1つであり、すでに持っていたことを除きます。もちろん、すべてのタスクを事前に確認する必要がありました。

コードは、Intel 64（x86-64）アーキテクチャのマルチコアハードウェアで実行されると暗黙的に想定されています。これを明示的に示す価値はあったかもしれませんが、追加の条件は誤解を招く可能性があります。さらに、問題の答えの説明が別の鉄の詳細を考慮に入れた場合、解決策がカウントされます。

タスク番号1

「Foreachは速い」、「Streamsは遅くなる」、「Parallelは速い」というのは、すでに4年前になったStream Java APIの議論に由来することがよくあります（Java 8のパブリックリリースの日付から数えると）だから私は挑発に屈した：

  void forEach(List<Integer> values, PrintStream ps) { values.forEach(ps::println); } void forEach(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } void forEach(List<Integer> values, PrintStream ps) { values.parallelStream().forEach(ps::println); }

List

明示的な実装、たとえばArrayList

を指定しないのは間違いでした。会議の参加者は、リストのどの実装が問題であるかを明確にするためにアプローチしました。すべてのベンチマークは、 ArrayList

とLinkedList

2つの標準実装に対して実施されたため、これらの条件で問題を解決することが提案されました。

明らかな間違った答え：オプション3は最速です。

多くの人が「並列という意味の速い」トリガーをトリガーし、 PrintStream

トラップされました。

PrintStreamソースコードスニペット：

  public void println(Object x) { String s = String.valueOf(x); synchronized (this) { print(s); newLine(); } }

実際には、これは、 values

リストのサイズに関係なく、並列実行のメリットを得られないことを意味しvalues

。それどころか、仮想マシンはPrintStream

複数のロックを常に解決する必要があるため、パフォーマンスが低下します。

`List.stream().forEach()`

vs `List.parallelStream().forEach()`

以下は、通常のシリアルとパラレルのStream

を比較するためのベンチマークです。

ベンチマーク

 package ru.gnkoshelev.jbreak2018.perf_tests.for_each; @Fork(value = 1, warmups = 0) @Warmup(iterations = 5, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @Measurement(iterations = 10, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @OutputTimeUnit(value = TimeUnit.MICROSECONDS) @BenchmarkMode(Mode.AverageTime) public class StreamDefaultBenchmark { static int N = 1000; static List<Integer> values; static { Random rand = new Random(12345); int size = N; values = new ArrayList<>(); for (int i = 0; i < size; i++) { values.add(rand.nextInt()); } } @State(value = Scope.Benchmark) public static class PrintStreamHolder { PrintStream ps; @Setup(value = Level.Iteration) public void setup() { ps = new PrintStream(new NullOutputStream()); } } @Benchmark public void forEachStreamBenchmark(PrintStreamHolder psh) { forEachStream(values, psh.ps); } @Benchmark public void forEachParallelStreamBenchmark(PrintStreamHolder psh) { forEachParallelStream(values, psh.ps); } public void forEachStream(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } public void forEachParallelStream(List<Integer> values, PrintStream ps) { values.parallelStream().forEach(ps::println); } }

結果

 # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamDefaultBenchmark.forEachParallelStreamBenchmark # Run progress: 0,00% complete, ETA 00:01:00 # Fork: 1 of 1 # Warmup Iteration 1: 905,302 us/op # Warmup Iteration 2: 876,525 us/op # Warmup Iteration 3: 921,153 us/op # Warmup Iteration 4: 898,899 us/op # Warmup Iteration 5: 873,496 us/op Iteration 1: 920,439 us/op Iteration 2: 894,216 us/op Iteration 3: 917,930 us/op Iteration 4: 906,970 us/op Iteration 5: 929,685 us/op Iteration 6: 883,136 us/op Iteration 7: 883,996 us/op Iteration 8: 882,597 us/op Iteration 9: 921,612 us/op Iteration 10: 885,576 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamDefaultBenchmark.forEachParallelStreamBenchmark": 902,616 ±(99.9%) 28,296 us/op [Average] (min, avg, max) = (882,597, 902,616, 929,685), stdev = 18,716 CI (99.9%): [874,320, 930,911] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamDefaultBenchmark.forEachStreamBenchmark # Run progress: 50,00% complete, ETA 00:00:30 # Fork: 1 of 1 # Warmup Iteration 1: 285,837 us/op # Warmup Iteration 2: 265,208 us/op # Warmup Iteration 3: 157,321 us/op # Warmup Iteration 4: 157,447 us/op # Warmup Iteration 5: 157,689 us/op Iteration 1: 157,192 us/op Iteration 2: 161,511 us/op Iteration 3: 161,464 us/op Iteration 4: 156,948 us/op Iteration 5: 158,526 us/op Iteration 6: 163,035 us/op Iteration 7: 159,140 us/op Iteration 8: 158,476 us/op Iteration 9: 158,884 us/op Iteration 10: 159,072 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamDefaultBenchmark.forEachStreamBenchmark": 159,425 ±(99.9%) 2,976 us/op [Average] (min, avg, max) = (156,948, 159,425, 163,035), stdev = 1,969 CI (99.9%): [156,448, 162,401] (assumes normal distribution) # Run complete. Total time: 00:01:00 Benchmark Mode Cnt Score Error Units StreamDefaultBenchmark.forEachParallelStreamBenchmark avgt 10 902,616 ± 28,296 us/op StreamDefaultBenchmark.forEachStreamBenchmark avgt 10 159,425 ± 2,976 us/op

シングルスレッド実行では、ロックがどこにも行かないことは明らかです。ただし、仮想マシンは非常に巧妙に動作することができます。つまり、（OSレベルで）実際のロックを行わず、リソースに競合がない場合、いわゆるバイアスロックに制限されます。

バイアスロックを無効にして、何が起こるかを確認できます。

バイアスロックなしのベンチマーク

 package ru.gnkoshelev.jbreak2018.perf_tests.for_each; @Fork(value = 1, warmups = 0, jvmArgsAppend = "-XX:-UseBiasedLocking") @Warmup(iterations = 5, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @Measurement(iterations = 10, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @OutputTimeUnit(value = TimeUnit.MICROSECONDS) @BenchmarkMode(Mode.AverageTime) public class StreamWithoutBiasedLockingBenchmark { static int N = 1000; static List<Integer> values; static { Random rand = new Random(12345); int size = N; values = new ArrayList<>(); for (int i = 0; i < size; i++) { values.add(rand.nextInt()); } } @State(value = Scope.Benchmark) public static class PrintStreamHolder { PrintStream ps; @Setup(value = Level.Iteration) public void setup() { ps = new PrintStream(new NullOutputStream()); } } @Benchmark public void forEachStreamBenchmark(PrintStreamHolder psh) { forEachStream(values, psh.ps); } @Benchmark public void forEachParallelStreamBenchmark(PrintStreamHolder psh) { forEachParallelStream(values, psh.ps); } public void forEachStream(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } public void forEachParallelStream(List<Integer> values, PrintStream ps) { values.parallelStream().forEach(ps::println); } }

結果

 # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: -XX:-UseBiasedLocking # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithoutBiasedLockingBenchmark.forEachParallelStreamBenchmark # Run progress: 0,00% complete, ETA 00:01:00 # Fork: 1 of 1 # Warmup Iteration 1: 754,310 us/op # Warmup Iteration 2: 723,277 us/op # Warmup Iteration 3: 682,845 us/op # Warmup Iteration 4: 696,635 us/op # Warmup Iteration 5: 690,811 us/op Iteration 1: 702,129 us/op Iteration 2: 729,542 us/op Iteration 3: 689,514 us/op Iteration 4: 716,482 us/op Iteration 5: 734,766 us/op Iteration 6: 684,455 us/op Iteration 7: 682,483 us/op Iteration 8: 706,857 us/op Iteration 9: 690,011 us/op Iteration 10: 694,427 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithoutBiasedLockingBenchmark.forEachParallelStreamBenchmark": 703,067 ±(99.9%) 28,058 us/op [Average] (min, avg, max) = (682,483, 703,067, 734,766), stdev = 18,559 CI (99.9%): [675,008, 731,125] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: -XX:-UseBiasedLocking # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithoutBiasedLockingBenchmark.forEachStreamBenchmark # Run progress: 50,00% complete, ETA 00:00:30 # Fork: 1 of 1 # Warmup Iteration 1: 271,938 us/op # Warmup Iteration 2: 258,261 us/op # Warmup Iteration 3: 257,976 us/op # Warmup Iteration 4: 256,103 us/op # Warmup Iteration 5: 255,863 us/op Iteration 1: 266,376 us/op Iteration 2: 258,158 us/op Iteration 3: 278,038 us/op Iteration 4: 271,354 us/op Iteration 5: 256,021 us/op Iteration 6: 254,590 us/op Iteration 7: 254,944 us/op Iteration 8: 255,525 us/op Iteration 9: 256,339 us/op Iteration 10: 257,311 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithoutBiasedLockingBenchmark.forEachStreamBenchmark": 260,866 ±(99.9%) 12,366 us/op [Average] (min, avg, max) = (254,590, 260,866, 278,038), stdev = 8,180 CI (99.9%): [248,499, 273,232] (assumes normal distribution) # Run complete. Total time: 00:01:00 Benchmark Mode Cnt Score Error Units StreamWithoutBiasedLockingBenchmark.forEachParallelStreamBenchmark avgt 10 703,067 ± 28,058 us/op StreamWithoutBiasedLockingBenchmark.forEachStreamBenchmark avgt 10 260,866 ± 12,366 us/op

予想通り、通常のArrayList.forEach()

のパフォーマンスを低下させました。同時に、3回目の反復から始まる加速は、前のベンチマークの温暖化の反復で顕著です-これは偶然ではありません。実際、デフォルトでは、仮想マシンが起動してから4000ミリ秒後にバイアスロックが有効になっています。この値を調整して、何が起こるかを確認できます。

オンタイムバイアスロックが強化されたベンチマーク

 package ru.gnkoshelev.jbreak2018.perf_tests.for_each; @Fork(value = 1, warmups = 0, jvmArgsAppend = "-XX:BiasedLockingStartupDelay=6000") @Warmup(iterations = 5, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @Measurement(iterations = 10, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @OutputTimeUnit(value = TimeUnit.MICROSECONDS) @BenchmarkMode(Mode.AverageTime) public class StreamWithNewBiasedLockingStartupDelayBenchmark { static int N = 1000; static List<Integer> values; static { Random rand = new Random(12345); int size = N; values = new ArrayList<>(); for (int i = 0; i < size; i++) { values.add(rand.nextInt()); } } @State(value = Scope.Benchmark) public static class PrintStreamHolder { PrintStream ps; @Setup(value = Level.Iteration) public void setup() { ps = new PrintStream(new NullOutputStream()); } } @Benchmark public void forEachStreamBenchmark(PrintStreamHolder psh) { forEachStream(values, psh.ps); } @Benchmark public void forEachParallelStreamBenchmark(PrintStreamHolder psh) { forEachParallelStream(values, psh.ps); } public void forEachStream(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } public void forEachParallelStream(List<Integer> values, PrintStream ps) { values.parallelStream().forEach(ps::println); } }

結果

 # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: -XX:BiasedLockingStartupDelay=6000 # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithNewBiasedLockingStartupDelayBenchmark.forEachParallelStreamBenchmark # Run progress: 0,00% complete, ETA 00:01:00 # Fork: 1 of 1 # Warmup Iteration 1: 954,533 us/op # Warmup Iteration 2: 866,854 us/op # Warmup Iteration 3: 907,109 us/op # Warmup Iteration 4: 914,717 us/op # Warmup Iteration 5: 924,102 us/op Iteration 1: 912,619 us/op Iteration 2: 947,812 us/op Iteration 3: 925,730 us/op Iteration 4: 933,807 us/op Iteration 5: 935,927 us/op Iteration 6: 852,369 us/op Iteration 7: 882,498 us/op Iteration 8: 852,625 us/op Iteration 9: 898,787 us/op Iteration 10: 1150,831 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithNewBiasedLockingStartupDelayBenchmark.forEachParallelStreamBenchmark": 929,301 ±(99.9%) 128,179 us/op [Average] (min, avg, max) = (852,369, 929,301, 1150,831), stdev = 84,783 CI (99.9%): [801,121, 1057,480] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: -XX:BiasedLockingStartupDelay=6000 # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithNewBiasedLockingStartupDelayBenchmark.forEachStreamBenchmark # Run progress: 50,00% complete, ETA 00:00:30 # Fork: 1 of 1 # Warmup Iteration 1: 346,853 us/op # Warmup Iteration 2: 280,643 us/op # Warmup Iteration 3: 264,425 us/op # Warmup Iteration 4: 169,200 us/op # Warmup Iteration 5: 167,981 us/op Iteration 1: 171,230 us/op Iteration 2: 169,891 us/op Iteration 3: 169,124 us/op Iteration 4: 167,938 us/op Iteration 5: 167,471 us/op Iteration 6: 176,187 us/op Iteration 7: 171,791 us/op Iteration 8: 170,127 us/op Iteration 9: 169,563 us/op Iteration 10: 169,062 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.StreamWithNewBiasedLockingStartupDelayBenchmark.forEachStreamBenchmark": 170,238 ±(99.9%) 3,736 us/op [Average] (min, avg, max) = (167,471, 170,238, 176,187), stdev = 2,471 CI (99.9%): [166,503, 173,974] (assumes normal distribution) # Run complete. Total time: 00:01:00 Benchmark Mode Cnt Score Error Units StreamWithNewBiasedLockingStartupDelayBenchmark.forEachParallelStreamBenchmark avgt 10 929,301 ± 128,179 us/op StreamWithNewBiasedLockingStartupDelayBenchmark.forEachStreamBenchmark avgt 10 170,238 ± 3,736 us/op

予想どおり、 バイアスロックの効果は、ベンチマークの開始から6秒後にのみ発生しました。これは、3回目と4回目のウォーミングイテレーションの時間の差に見られます。

List.forEach（）vs List.stream（）。ForEach（）

次に、どちらが速いかをList.stream().forEach()

ます： List.forEach()

またはList.stream().forEach()

。正解は「に依存する」です。結果は、リストのサイズ、リストのタイプ、インライン化を使用する能力、およびJITコンパイラーによるその他の最適化の影響を受けます。一般的なルールを特定できる場合があります。 ArrayList

の結果は次のとおりArrayList

。

ベンチマーク

 package ru.gnkoshelev.jbreak2018.perf_tests.for_each; @Fork(value = 1, warmups = 0) @Warmup(iterations = 5, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @Measurement(iterations = 10, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @OutputTimeUnit(value = TimeUnit.MICROSECONDS) @BenchmarkMode(Mode.AverageTime) @State(Scope.Benchmark) public class ArrayListVsStreamBenchmark { @Param(value = {"1", "10", "100", "1000", "10000"}) public int N; private List<Integer> values; @Setup public void setup() { Random rand = new Random(12345); int size = N; values = new ArrayList<>(); for (int i = 0; i < size; i++) { values.add(rand.nextInt()); } } @State(value = Scope.Benchmark) public static class PrintStreamHolder { PrintStream ps; @Setup(value = Level.Iteration) public void setup() { ps = new PrintStream(new NullOutputStream()); } } @Benchmark public void forEachListBenchmark(PrintStreamHolder psh) { forEachList(values, psh.ps); } @Benchmark public void forEachStreamBenchmark(PrintStreamHolder psh) { forEachStream(values, psh.ps); } public void forEachList(List<Integer> values, PrintStream ps) { values.forEach(ps::println); } public void forEachStream(List<Integer> values, PrintStream ps) { values.stream().forEach(ps::println); } }

結果

 # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark # Parameters: (N = 1) # Run progress: 0,00% complete, ETA 00:05:00 # Fork: 1 of 1 # Warmup Iteration 1: 0,288 us/op # Warmup Iteration 2: 0,277 us/op # Warmup Iteration 3: 0,178 us/op # Warmup Iteration 4: 0,177 us/op # Warmup Iteration 5: 0,178 us/op Iteration 1: 0,177 us/op Iteration 2: 0,178 us/op Iteration 3: 0,178 us/op Iteration 4: 0,178 us/op Iteration 5: 0,178 us/op Iteration 6: 0,178 us/op Iteration 7: 0,178 us/op Iteration 8: 0,179 us/op Iteration 9: 0,178 us/op Iteration 10: 0,180 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark": 0,178 ±(99.9%) 0,001 us/op [Average] (min, avg, max) = (0,177, 0,178, 0,180), stdev = 0,001 CI (99.9%): [0,177, 0,179] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark # Parameters: (N = 10) # Run progress: 10,00% complete, ETA 00:04:33 # Fork: 1 of 1 # Warmup Iteration 1: 2,933 us/op # Warmup Iteration 2: 2,839 us/op # Warmup Iteration 3: 1,661 us/op # Warmup Iteration 4: 1,675 us/op # Warmup Iteration 5: 1,674 us/op Iteration 1: 1,682 us/op Iteration 2: 1,653 us/op Iteration 3: 1,658 us/op Iteration 4: 1,656 us/op Iteration 5: 1,659 us/op Iteration 6: 1,655 us/op Iteration 7: 1,656 us/op Iteration 8: 1,657 us/op Iteration 9: 1,661 us/op Iteration 10: 1,660 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark": 1,660 ±(99.9%) 0,013 us/op [Average] (min, avg, max) = (1,653, 1,660, 1,682), stdev = 0,008 CI (99.9%): [1,647, 1,672] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark # Parameters: (N = 100) # Run progress: 20,00% complete, ETA 00:04:02 # Fork: 1 of 1 # Warmup Iteration 1: 27,633 us/op # Warmup Iteration 2: 27,184 us/op # Warmup Iteration 3: 15,046 us/op # Warmup Iteration 4: 15,064 us/op # Warmup Iteration 5: 15,060 us/op Iteration 1: 15,039 us/op Iteration 2: 15,057 us/op Iteration 3: 15,065 us/op Iteration 4: 15,062 us/op Iteration 5: 15,086 us/op Iteration 6: 15,060 us/op Iteration 7: 15,110 us/op Iteration 8: 15,070 us/op Iteration 9: 15,111 us/op Iteration 10: 15,079 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark": 15,074 ±(99.9%) 0,035 us/op [Average] (min, avg, max) = (15,039, 15,074, 15,111), stdev = 0,023 CI (99.9%): [15,039, 15,109] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark # Parameters: (N = 1000) # Run progress: 30,00% complete, ETA 00:03:32 # Fork: 1 of 1 # Warmup Iteration 1: 291,532 us/op # Warmup Iteration 2: 267,136 us/op # Warmup Iteration 3: 170,432 us/op # Warmup Iteration 4: 170,388 us/op # Warmup Iteration 5: 169,797 us/op Iteration 1: 170,036 us/op Iteration 2: 176,614 us/op Iteration 3: 176,396 us/op Iteration 4: 175,895 us/op Iteration 5: 176,984 us/op Iteration 6: 172,085 us/op Iteration 7: 170,193 us/op Iteration 8: 171,333 us/op Iteration 9: 170,293 us/op Iteration 10: 171,006 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark": 173,084 ±(99.9%) 4,518 us/op [Average] (min, avg, max) = (170,036, 173,084, 176,984), stdev = 2,988 CI (99.9%): [168,566, 177,601] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark # Parameters: (N = 10000) # Run progress: 40,00% complete, ETA 00:03:01 # Fork: 1 of 1 # Warmup Iteration 1: 2890,241 us/op # Warmup Iteration 2: 2784,740 us/op # Warmup Iteration 3: 1725,390 us/op # Warmup Iteration 4: 1726,138 us/op # Warmup Iteration 5: 1733,212 us/op Iteration 1: 1726,084 us/op Iteration 2: 1712,973 us/op Iteration 3: 1715,916 us/op Iteration 4: 1750,530 us/op Iteration 5: 1721,900 us/op Iteration 6: 1711,158 us/op Iteration 7: 1709,659 us/op Iteration 8: 1726,751 us/op Iteration 9: 1737,237 us/op Iteration 10: 1734,220 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachListBenchmark": 1724,643 ±(99.9%) 19,861 us/op [Average] (min, avg, max) = (1709,659, 1724,643, 1750,530), stdev = 13,137 CI (99.9%): [1704,782, 1744,504] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark # Parameters: (N = 1) # Run progress: 50,00% complete, ETA 00:02:31 # Fork: 1 of 1 # Warmup Iteration 1: 0,301 us/op # Warmup Iteration 2: 0,289 us/op # Warmup Iteration 3: 0,181 us/op # Warmup Iteration 4: 0,180 us/op # Warmup Iteration 5: 0,181 us/op Iteration 1: 0,179 us/op Iteration 2: 0,178 us/op Iteration 3: 0,179 us/op Iteration 4: 0,180 us/op Iteration 5: 0,179 us/op Iteration 6: 0,179 us/op Iteration 7: 0,180 us/op Iteration 8: 0,179 us/op Iteration 9: 0,180 us/op Iteration 10: 0,180 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark": 0,179 ±(99.9%) 0,001 us/op [Average] (min, avg, max) = (0,178, 0,179, 0,180), stdev = 0,001 CI (99.9%): [0,178, 0,180] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark # Parameters: (N = 10) # Run progress: 60,00% complete, ETA 00:02:01 # Fork: 1 of 1 # Warmup Iteration 1: 2,660 us/op # Warmup Iteration 2: 2,551 us/op # Warmup Iteration 3: 1,558 us/op # Warmup Iteration 4: 1,563 us/op # Warmup Iteration 5: 1,563 us/op Iteration 1: 1,557 us/op Iteration 2: 1,553 us/op Iteration 3: 1,567 us/op Iteration 4: 1,555 us/op Iteration 5: 1,569 us/op Iteration 6: 1,570 us/op Iteration 7: 1,562 us/op Iteration 8: 1,561 us/op Iteration 9: 1,564 us/op Iteration 10: 1,580 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark": 1,564 ±(99.9%) 0,012 us/op [Average] (min, avg, max) = (1,553, 1,564, 1,580), stdev = 0,008 CI (99.9%): [1,552, 1,576] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark # Parameters: (N = 100) # Run progress: 70,00% complete, ETA 00:01:30 # Fork: 1 of 1 # Warmup Iteration 1: 27,754 us/op # Warmup Iteration 2: 26,859 us/op # Warmup Iteration 3: 16,456 us/op # Warmup Iteration 4: 16,548 us/op # Warmup Iteration 5: 16,432 us/op Iteration 1: 16,399 us/op Iteration 2: 16,480 us/op Iteration 3: 16,517 us/op Iteration 4: 16,458 us/op Iteration 5: 16,395 us/op Iteration 6: 16,471 us/op Iteration 7: 16,595 us/op Iteration 8: 16,585 us/op Iteration 9: 16,460 us/op Iteration 10: 16,435 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark": 16,479 ±(99.9%) 0,104 us/op [Average] (min, avg, max) = (16,395, 16,479, 16,595), stdev = 0,069 CI (99.9%): [16,376, 16,583] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark # Parameters: (N = 1000) # Run progress: 80,00% complete, ETA 00:01:00 # Fork: 1 of 1 # Warmup Iteration 1: 291,695 us/op # Warmup Iteration 2: 282,896 us/op # Warmup Iteration 3: 185,443 us/op # Warmup Iteration 4: 187,851 us/op # Warmup Iteration 5: 184,393 us/op Iteration 1: 184,584 us/op Iteration 2: 185,349 us/op Iteration 3: 184,803 us/op Iteration 4: 184,394 us/op Iteration 5: 185,371 us/op Iteration 6: 186,735 us/op Iteration 7: 185,945 us/op Iteration 8: 188,592 us/op Iteration 9: 186,581 us/op Iteration 10: 187,908 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark": 186,026 ±(99.9%) 2,142 us/op [Average] (min, avg, max) = (184,394, 186,026, 188,592), stdev = 1,417 CI (99.9%): [183,884, 188,168] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark # Parameters: (N = 10000) # Run progress: 90,00% complete, ETA 00:00:30 # Fork: 1 of 1 # Warmup Iteration 1: 2896,624 us/op # Warmup Iteration 2: 2816,471 us/op # Warmup Iteration 3: 1732,597 us/op # Warmup Iteration 4: 1798,371 us/op # Warmup Iteration 5: 1758,489 us/op Iteration 1: 1615,213 us/op Iteration 2: 1518,388 us/op Iteration 3: 1513,955 us/op Iteration 4: 1520,570 us/op Iteration 5: 1525,072 us/op Iteration 6: 1527,055 us/op Iteration 7: 1547,707 us/op Iteration 8: 1532,163 us/op Iteration 9: 1519,474 us/op Iteration 10: 1529,969 us/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.ArrayListVsStreamBenchmark.forEachStreamBenchmark": 1534,957 ±(99.9%) 44,959 us/op [Average] (min, avg, max) = (1513,955, 1534,957, 1615,213), stdev = 29,737 CI (99.9%): [1489,998, 1579,915] (assumes normal distribution) # Run complete. Total time: 00:05:03 Benchmark (N) Mode Cnt Score Error Units ArrayListVsStreamBenchmark.forEachListBenchmark 1 avgt 10 0,178 ± 0,001 us/op ArrayListVsStreamBenchmark.forEachListBenchmark 10 avgt 10 1,660 ± 0,013 us/op ArrayListVsStreamBenchmark.forEachListBenchmark 100 avgt 10 15,074 ± 0,035 us/op ArrayListVsStreamBenchmark.forEachListBenchmark 1000 avgt 10 173,084 ± 4,518 us/op ArrayListVsStreamBenchmark.forEachListBenchmark 10000 avgt 10 1724,643 ± 19,861 us/op ArrayListVsStreamBenchmark.forEachStreamBenchmark 1 avgt 10 0,179 ± 0,001 us/op ArrayListVsStreamBenchmark.forEachStreamBenchmark 10 avgt 10 1,564 ± 0,012 us/op ArrayListVsStreamBenchmark.forEachStreamBenchmark 100 avgt 10 16,479 ± 0,104 us/op ArrayListVsStreamBenchmark.forEachStreamBenchmark 1000 avgt 10 186,026 ± 2,142 us/op ArrayListVsStreamBenchmark.forEachStreamBenchmark 10000 avgt 10 1534,957 ± 44,959 us/op

実際、反復のオーバーヘッドは、メソッドを呼び出すオーバーヘッドよりもはるかに小さくなりPrintStream.println()

ます。これを確認するには、メソッド呼び出しPrintStrim.println()

を呼び出しに置き換えるだけで十分ですBlackhole.consume()

（後者は、JITコンパイラーがコードの役に立たなくなった部分を切り取らないことを保証し、反復が機能します）。結果はマイクロ秒単位ではなく、ナノ秒単位であることに注意してください！

Blackhole.consume（）のベンチマーク

 package ru.gnkoshelev.jbreak2018.perf_tests.for_each; @Fork(value = 1, warmups = 0) @Warmup(iterations = 5, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @Measurement(iterations = 10, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @OutputTimeUnit(value = TimeUnit.NANOSECONDS) @BenchmarkMode(Mode.AverageTime) @State(Scope.Benchmark) public class BlackholeConsumingBenchmark { @Param(value = {"1", "10", "100", "1000", "10000"}) public int N; private List<Integer> values; @Setup public void setup() { Random rand = new Random(12345); int size = N; values = new ArrayList<>(); for (int i = 0; i < size; i++) { values.add(rand.nextInt()); } } @Benchmark public void forEachListBenchmark(Blackhole bh) { forEachList(values, bh); } @Benchmark public void forEachStreamBenchmark(Blackhole bh) { forEachStream(values, bh); } public void forEachList(List<Integer> values, Blackhole bh) { values.forEach(bh::consume); } public void forEachStream(List<Integer> values, Blackhole bh) { values.stream().forEach(bh::consume); } }

結果

 # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark # Parameters: (N = 1) # Run progress: 0,00% complete, ETA 00:05:00 # Fork: 1 of 1 # Warmup Iteration 1: 6,866 ns/op # Warmup Iteration 2: 6,865 ns/op # Warmup Iteration 3: 6,528 ns/op # Warmup Iteration 4: 6,524 ns/op # Warmup Iteration 5: 6,521 ns/op Iteration 1: 6,519 ns/op Iteration 2: 6,542 ns/op Iteration 3: 6,608 ns/op Iteration 4: 6,590 ns/op Iteration 5: 6,678 ns/op Iteration 6: 6,658 ns/op Iteration 7: 6,756 ns/op Iteration 8: 6,933 ns/op Iteration 9: 6,939 ns/op Iteration 10: 6,811 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark": 6,703 ±(99.9%) 0,230 ns/op [Average] (min, avg, max) = (6,519, 6,703, 6,939), stdev = 0,152 CI (99.9%): [6,474, 6,933] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark # Parameters: (N = 10) # Run progress: 10,00% complete, ETA 00:04:32 # Fork: 1 of 1 # Warmup Iteration 1: 51,655 ns/op # Warmup Iteration 2: 50,507 ns/op # Warmup Iteration 3: 50,309 ns/op # Warmup Iteration 4: 50,886 ns/op # Warmup Iteration 5: 50,068 ns/op Iteration 1: 51,218 ns/op Iteration 2: 50,108 ns/op Iteration 3: 52,343 ns/op Iteration 4: 50,745 ns/op Iteration 5: 51,461 ns/op Iteration 6: 50,366 ns/op Iteration 7: 49,976 ns/op Iteration 8: 50,623 ns/op Iteration 9: 50,223 ns/op Iteration 10: 50,125 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark": 50,719 ±(99.9%) 1,138 ns/op [Average] (min, avg, max) = (49,976, 50,719, 52,343), stdev = 0,753 CI (99.9%): [49,581, 51,857] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark # Parameters: (N = 100) # Run progress: 20,00% complete, ETA 00:04:02 # Fork: 1 of 1 # Warmup Iteration 1: 574,874 ns/op # Warmup Iteration 2: 557,543 ns/op # Warmup Iteration 3: 557,260 ns/op # Warmup Iteration 4: 556,273 ns/op # Warmup Iteration 5: 556,025 ns/op Iteration 1: 555,532 ns/op Iteration 2: 558,421 ns/op Iteration 3: 564,582 ns/op Iteration 4: 573,368 ns/op Iteration 5: 576,355 ns/op Iteration 6: 562,087 ns/op Iteration 7: 565,229 ns/op Iteration 8: 574,557 ns/op Iteration 9: 561,430 ns/op Iteration 10: 574,486 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark": 566,605 ±(99.9%) 11,369 ns/op [Average] (min, avg, max) = (555,532, 566,605, 576,355), stdev = 7,520 CI (99.9%): [555,236, 577,974] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark # Parameters: (N = 1000) # Run progress: 30,00% complete, ETA 00:03:32 # Fork: 1 of 1 # Warmup Iteration 1: 4782,869 ns/op # Warmup Iteration 2: 4920,389 ns/op # Warmup Iteration 3: 4845,857 ns/op # Warmup Iteration 4: 4847,470 ns/op # Warmup Iteration 5: 4857,726 ns/op Iteration 1: 4850,519 ns/op Iteration 2: 4852,075 ns/op Iteration 3: 4907,015 ns/op Iteration 4: 4859,020 ns/op Iteration 5: 4879,201 ns/op Iteration 6: 4832,822 ns/op Iteration 7: 4831,532 ns/op Iteration 8: 4829,599 ns/op Iteration 9: 4878,201 ns/op Iteration 10: 4826,464 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark": 4854,645 ±(99.9%) 40,238 ns/op [Average] (min, avg, max) = (4826,464, 4854,645, 4907,015), stdev = 26,615 CI (99.9%): [4814,407, 4894,883] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark # Parameters: (N = 10000) # Run progress: 40,00% complete, ETA 00:03:01 # Fork: 1 of 1 # Warmup Iteration 1: 42925,340 ns/op # Warmup Iteration 2: 47991,419 ns/op # Warmup Iteration 3: 50854,318 ns/op # Warmup Iteration 4: 51251,774 ns/op # Warmup Iteration 5: 51756,902 ns/op Iteration 1: 51102,960 ns/op Iteration 2: 51420,802 ns/op Iteration 3: 51744,142 ns/op Iteration 4: 50985,401 ns/op Iteration 5: 51125,197 ns/op Iteration 6: 51223,229 ns/op Iteration 7: 51424,190 ns/op Iteration 8: 51457,326 ns/op Iteration 9: 51410,258 ns/op Iteration 10: 51419,014 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachListBenchmark": 51331,252 ±(99.9%) 335,358 ns/op [Average] (min, avg, max) = (50985,401, 51331,252, 51744,142), stdev = 221,819 CI (99.9%): [50995,894, 51666,610] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 1) # Run progress: 50,00% complete, ETA 00:02:31 # Fork: 1 of 1 # Warmup Iteration 1: 25,963 ns/op # Warmup Iteration 2: 20,840 ns/op # Warmup Iteration 3: 18,882 ns/op # Warmup Iteration 4: 19,081 ns/op # Warmup Iteration 5: 19,513 ns/op Iteration 1: 19,411 ns/op Iteration 2: 18,856 ns/op Iteration 3: 19,049 ns/op Iteration 4: 18,891 ns/op Iteration 5: 19,136 ns/op Iteration 6: 20,116 ns/op Iteration 7: 19,490 ns/op Iteration 8: 19,251 ns/op Iteration 9: 19,767 ns/op Iteration 10: 19,607 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark": 19,357 ±(99.9%) 0,607 ns/op [Average] (min, avg, max) = (18,856, 19,357, 20,116), stdev = 0,401 CI (99.9%): [18,751, 19,964] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 10) # Run progress: 60,00% complete, ETA 00:02:01 # Fork: 1 of 1 # Warmup Iteration 1: 59,828 ns/op # Warmup Iteration 2: 52,901 ns/op # Warmup Iteration 3: 44,941 ns/op # Warmup Iteration 4: 44,833 ns/op # Warmup Iteration 5: 44,962 ns/op Iteration 1: 46,032 ns/op Iteration 2: 45,802 ns/op Iteration 3: 45,018 ns/op Iteration 4: 45,704 ns/op Iteration 5: 45,788 ns/op Iteration 6: 46,674 ns/op Iteration 7: 46,588 ns/op Iteration 8: 48,083 ns/op Iteration 9: 46,862 ns/op Iteration 10: 47,087 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark": 46,364 ±(99.9%) 1,321 ns/op [Average] (min, avg, max) = (45,018, 46,364, 48,083), stdev = 0,874 CI (99.9%): [45,042, 47,685] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 100) # Run progress: 70,00% complete, ETA 00:01:31 # Fork: 1 of 1 # Warmup Iteration 1: 382,038 ns/op # Warmup Iteration 2: 347,846 ns/op # Warmup Iteration 3: 337,712 ns/op # Warmup Iteration 4: 335,078 ns/op # Warmup Iteration 5: 344,895 ns/op Iteration 1: 329,577 ns/op Iteration 2: 328,436 ns/op Iteration 3: 329,517 ns/op Iteration 4: 330,736 ns/op Iteration 5: 335,464 ns/op Iteration 6: 338,727 ns/op Iteration 7: 330,996 ns/op Iteration 8: 330,809 ns/op Iteration 9: 329,599 ns/op Iteration 10: 330,580 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark": 331,444 ±(99.9%) 4,800 ns/op [Average] (min, avg, max) = (328,436, 331,444, 338,727), stdev = 3,175 CI (99.9%): [326,645, 336,244] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 1000) # Run progress: 80,00% complete, ETA 00:01:00 # Fork: 1 of 1 # Warmup Iteration 1: 3734,819 ns/op # Warmup Iteration 2: 3534,547 ns/op # Warmup Iteration 3: 3201,812 ns/op # Warmup Iteration 4: 3151,474 ns/op # Warmup Iteration 5: 3136,897 ns/op Iteration 1: 3138,505 ns/op Iteration 2: 3137,993 ns/op Iteration 3: 3141,723 ns/op Iteration 4: 3225,711 ns/op Iteration 5: 3127,945 ns/op Iteration 6: 3129,312 ns/op Iteration 7: 3141,355 ns/op Iteration 8: 3137,377 ns/op Iteration 9: 3149,419 ns/op Iteration 10: 3243,662 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark": 3157,300 ±(99.9%) 62,671 ns/op [Average] (min, avg, max) = (3127,945, 3157,300, 3243,662), stdev = 41,453 CI (99.9%): [3094,629, 3219,972] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 10000) # Run progress: 90,00% complete, ETA 00:00:30 # Fork: 1 of 1 # Warmup Iteration 1: 45988,758 ns/op # Warmup Iteration 2: 45331,295 ns/op # Warmup Iteration 3: 38213,299 ns/op # Warmup Iteration 4: 31233,461 ns/op # Warmup Iteration 5: 31321,451 ns/op Iteration 1: 31470,793 ns/op Iteration 2: 32267,081 ns/op Iteration 3: 31619,772 ns/op Iteration 4: 32155,094 ns/op Iteration 5: 31304,547 ns/op Iteration 6: 31484,244 ns/op Iteration 7: 31709,801 ns/op Iteration 8: 31297,118 ns/op Iteration 9: 31305,782 ns/op Iteration 10: 31320,821 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.BlackholeConsumingBenchmark.forEachStreamBenchmark": 31593,505 ±(99.9%) 537,786 ns/op [Average] (min, avg, max) = (31297,118, 31593,505, 32267,081), stdev = 355,712 CI (99.9%): [31055,719, 32131,292] (assumes normal distribution) # Run complete. Total time: 00:05:03 Benchmark (N) Mode Cnt Score Error Units BlackholeConsumingBenchmark.forEachListBenchmark 1 avgt 10 6,703 ± 0,230 ns/op BlackholeConsumingBenchmark.forEachListBenchmark 10 avgt 10 50,719 ± 1,138 ns/op BlackholeConsumingBenchmark.forEachListBenchmark 100 avgt 10 566,605 ± 11,369 ns/op BlackholeConsumingBenchmark.forEachListBenchmark 1000 avgt 10 4854,645 ± 40,238 ns/op BlackholeConsumingBenchmark.forEachListBenchmark 10000 avgt 10 51331,252 ± 335,358 ns/op BlackholeConsumingBenchmark.forEachStreamBenchmark 1 avgt 10 19,357 ± 0,607 ns/op BlackholeConsumingBenchmark.forEachStreamBenchmark 10 avgt 10 46,364 ± 1,321 ns/op BlackholeConsumingBenchmark.forEachStreamBenchmark 100 avgt 10 331,444 ± 4,800 ns/op BlackholeConsumingBenchmark.forEachStreamBenchmark 1000 avgt 10 3157,300 ± 62,671 ns/op BlackholeConsumingBenchmark.forEachStreamBenchmark 10000 avgt 10 31593,505 ± 537,786 ns/op

そして、それはより興味深いものになります：との違いはArrayList.forEach()

、ArrayList.stream().forEach()

長いリストで非常に大きくなりました。さらに、加熱は著しく加速しますがArrayList.forEach()

、加速することはありませんArrayList.stream().forEach()

。

実際、作成のオーバーヘッドStream

は小さなリストにのみ影響します。将来的には、直接的な反復になります。

次のようになりますArrayList.forEach()

。

  public void forEach(Consumer<? super E> action) { Objects.requireNonNull(action); final int expectedModCount = modCount; @SuppressWarnings("unchecked") final E[] elementData = (E[]) this.elementData; final int size = this.size; for (int i=0; modCount == expectedModCount && i < size; i++) { action.accept(elementData[i]); } if (modCount != expectedModCount) { throw new ConcurrentModificationException(); } }

コードは次のようになりArrayList.ArrayListSpliterator.forEachRemaining()

ます。

  public void forEachRemaining(Consumer<? super E> action) { int i, hi, mc; // hoist accesses and checks from loop ArrayList<E> lst; Object[] a; if (action == null) throw new NullPointerException(); if ((lst = list) != null && (a = lst.elementData) != null) { if ((hi = fence) < 0) { mc = lst.modCount; hi = lst.size; } else mc = expectedModCount; if ((i = index) >= 0 && (index = hi) <= a.length) { for (; i < hi; ++i) { @SuppressWarnings("unchecked") E e = (E) a[i]; action.accept(e); } if (lst.modCount == mc) return; } } throw new ConcurrentModificationException(); }

ご覧のとおり、コードの違いはそれほど重要ではありません（メソッド内のループの場合）。JITコンパイラーが作用します。Andrei apanginは、何が起こっているのかを把握するのに役立ちました。Blackhole.consume()

インライン化されず、JITコンパイラーはコードを適切に最適化できません。その結果、フィールドを読み取りmodCount

、配列の境界をチェックします。

実際、私たちはそれBlackhole.consume()

を私たちの自作の方法に置き換えようとします：

価値の別の消費者とのベンチマーク

 package ru.gnkoshelev.jbreak2018.perf_tests.for_each; @Fork(value = 1, warmups = 0) @Warmup(iterations = 5, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @Measurement(iterations = 10, time = 2_000, timeUnit = TimeUnit.MILLISECONDS) @OutputTimeUnit(value = TimeUnit.NANOSECONDS) @BenchmarkMode(Mode.AverageTime) @State(Scope.Benchmark) public class TrickyConsumingBenchmark { @Param(value = {"100", "10000"}) public int N; private List<Integer> values; @Setup public void setup() { Random rand = new Random(12345); int size = N; values = new ArrayList<>(); for (int i = 0; i < size; i++) { values.add(rand.nextInt()); } } @Benchmark public void forEachListBenchmark() { forEachList(values); } @Benchmark public void forEachStreamBenchmark() { forEachStream(values); } public void forEachList(List<Integer> values) { values.forEach(TrickyConsumer::consume); } public void forEachStream(List<Integer> values) { values.stream().forEach(TrickyConsumer::consume); } public static class TrickyConsumer { public static Integer value; public static void consume(Integer value) { if (! (value % 100 == 0)) { TrickyConsumer.value = value; } } } }

結果

 # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachListBenchmark # Parameters: (N = 100) # Run progress: 0,00% complete, ETA 00:02:00 # Fork: 1 of 1 # Warmup Iteration 1: 205,489 ns/op # Warmup Iteration 2: 205,514 ns/op # Warmup Iteration 3: 216,056 ns/op # Warmup Iteration 4: 208,956 ns/op # Warmup Iteration 5: 204,900 ns/op Iteration 1: 205,535 ns/op Iteration 2: 210,834 ns/op Iteration 3: 206,050 ns/op Iteration 4: 205,962 ns/op Iteration 5: 205,285 ns/op Iteration 6: 205,581 ns/op Iteration 7: 204,991 ns/op Iteration 8: 209,290 ns/op Iteration 9: 203,927 ns/op Iteration 10: 204,977 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachListBenchmark": 206,243 ±(99.9%) 3,222 ns/op [Average] (min, avg, max) = (203,927, 206,243, 210,834), stdev = 2,131 CI (99.9%): [203,021, 209,465] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachListBenchmark # Parameters: (N = 10000) # Run progress: 25,00% complete, ETA 00:01:30 # Fork: 1 of 1 # Warmup Iteration 1: 22733,300 ns/op # Warmup Iteration 2: 22197,064 ns/op # Warmup Iteration 3: 21640,616 ns/op # Warmup Iteration 4: 21411,160 ns/op # Warmup Iteration 5: 21702,499 ns/op Iteration 1: 21921,831 ns/op Iteration 2: 21970,321 ns/op Iteration 3: 21409,419 ns/op Iteration 4: 21171,131 ns/op Iteration 5: 22011,961 ns/op Iteration 6: 22063,926 ns/op Iteration 7: 21825,015 ns/op Iteration 8: 21820,783 ns/op Iteration 9: 22089,657 ns/op Iteration 10: 21626,934 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachListBenchmark": 21791,098 ±(99.9%) 456,511 ns/op [Average] (min, avg, max) = (21171,131, 21791,098, 22089,657), stdev = 301,954 CI (99.9%): [21334,587, 22247,609] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 100) # Run progress: 50,00% complete, ETA 00:01:00 # Fork: 1 of 1 # Warmup Iteration 1: 250,480 ns/op # Warmup Iteration 2: 247,070 ns/op # Warmup Iteration 3: 269,745 ns/op # Warmup Iteration 4: 269,914 ns/op # Warmup Iteration 5: 272,912 ns/op Iteration 1: 269,779 ns/op Iteration 2: 270,502 ns/op Iteration 3: 269,128 ns/op Iteration 4: 273,862 ns/op Iteration 5: 275,447 ns/op Iteration 6: 272,090 ns/op Iteration 7: 271,189 ns/op Iteration 8: 271,714 ns/op Iteration 9: 269,121 ns/op Iteration 10: 269,913 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachStreamBenchmark": 271,275 ±(99.9%) 3,144 ns/op [Average] (min, avg, max) = (269,121, 271,275, 275,447), stdev = 2,079 CI (99.9%): [268,131, 274,418] (assumes normal distribution) # JMH version: 1.20 # VM version: JDK 1.8.0_161, VM 25.161-b12 # VM invoker: C:\Program Files\Java\jre1.8.0_161\bin\java.exe # VM options: <none> # Warmup: 5 iterations, 2000 ms each # Measurement: 10 iterations, 2000 ms each # Timeout: 10 min per iteration # Threads: 1 thread, will synchronize iterations # Benchmark mode: Average time, time/op # Benchmark: ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachStreamBenchmark # Parameters: (N = 10000) # Run progress: 75,00% complete, ETA 00:00:30 # Fork: 1 of 1 # Warmup Iteration 1: 22612,079 ns/op # Warmup Iteration 2: 24943,260 ns/op # Warmup Iteration 3: 23366,453 ns/op # Warmup Iteration 4: 23445,251 ns/op # Warmup Iteration 5: 23732,635 ns/op Iteration 1: 23575,011 ns/op Iteration 2: 23478,581 ns/op Iteration 3: 23663,154 ns/op Iteration 4: 23067,535 ns/op Iteration 5: 23489,020 ns/op Iteration 6: 23461,241 ns/op Iteration 7: 23510,542 ns/op Iteration 8: 23504,541 ns/op Iteration 9: 24036,074 ns/op Iteration 10: 24081,512 ns/op Result "ru.gnkoshelev.jbreak2018.perf_tests.for_each.TrickyConsumingBenchmark.forEachStreamBenchmark": 23586,721 ±(99.9%) 442,736 ns/op [Average] (min, avg, max) = (23067,535, 23586,721, 24081,512), stdev = 292,843 CI (99.9%): [23143,985, 24029,458] (assumes normal distribution) # Run complete. Total time: 00:02:01 Benchmark (N) Mode Cnt Score Error Units TrickyConsumingBenchmark.forEachListBenchmark 100 avgt 10 206,243 ± 3,222 ns/op TrickyConsumingBenchmark.forEachListBenchmark 10000 avgt 10 21791,098 ± 456,511 ns/op TrickyConsumingBenchmark.forEachStreamBenchmark 100 avgt 10 271,275 ± 3,144 ns/op TrickyConsumingBenchmark.forEachStreamBenchmark 10000 avgt 10 23586,721 ± 442,736 ns/op

すべてが正常に戻り、ここでJITがどのように機能したかはまったく別の話です。

結論

1.最終的に正しい答えは何ですか？問題の方法1と2の結果はわずかに異なり、方法3はパフォーマンスが著しく劣ります。

2.別の選択肢が正解として受け入れられました。1が2よりも速く、2が正しく議論された場合は1よりも速いということです。

3.個人的な結論：ウサギの穴は、あなたが思っているよりも常に深い。

ベンチマークコードはgithubで取得できます：jbreak2018-forEach-perf-tests。

統計

提出された32のオプションのうち、8つが正しく、5つが部分的に正しかった。別の7人が、最速PrintStream

と並行しStream

て、consideringに陥りました。連続したストリームを信じている人はおらず、オプション2を最速として選択しませんでした。

加算

以下のコメントList

では、会議が手動で防御された任意の実装に関連する0日間の脆弱性を使用した潜在的な悪用についてはすでに議論されています。Tagir lanyに

感謝します。彼はそのようなエクスプロイトのバリアントを実装しました（彼の記事では、エクスプロイトコード、詳細な説明と例）。

PS

来週末、次のパートを準備しようと思います。

UPD。シリーズの他の出版物：パート2、パート3、パート4。

JBreakを使用したパフォーマンスタスクの分析（パート1）

まえがき

タスク

タスク番号1

List.forEach（）vs List.stream（）。ForEach（）

結論

統計

加算

PS

More articles: