在查看一些分析结果时，我注意到在一个紧密循环中使用流（而不是另一个嵌套的循环）会产生大量类型为java.util.stream.ReferencePipeline和java.util.ArrayList$ArrayListSpliterator的对象的内存开销。我将有问题的流转换为for-each循环后，内存消耗显著降低。
我知道流不能保证比普通循环执行得更好，但我本以为差别微不足道。在这种情况下，似乎增加了40%。
以下是我编写的测试类，用于隔离问题。我使用JFR监视内存消耗和对象分配：
import java.util.ArrayList;
import java.util.List;
import java.util.Optional;
import java.util.Random;
import java.util.function.Predicate;
public class StreamMemoryTest {
private static boolean blackHole = false;
public static List<Integer> getRandListOfSize(int size) {
ArrayList<Integer> randList = new ArrayList<>(size);
Random rnGen = new Random();
for (int i = 0; i < size; i++) {
randList.add(rnGen.nextInt(100));
}
return randList;
}
public static boolean getIndexOfNothingManualImpl(List<Integer> nums, Predicate<Integer> predicate) {
for (Integer num : nums) {
// Impossible condition
if (predicate.test(num)) {
return true;
}
}
return false;
}
public static boolean getIndexOfNothingStreamImpl(List<Integer> nums, Predicate<Integer> predicate) {
Optional<Integer> first = nums.stream().filter(predicate).findFirst();
return first.isPresent();
}
public static void consume(boolean value) {
blackHole = blackHole && value;
}
public static boolean result() {
return blackHole;
}
public static void main(String[] args) {
// 100 million trials
int numTrials = 100000000;
System.out.println("Beginning test");
for (int i = 0; i < numTrials; i++) {
List<Integer> randomNums = StreamMemoryTest.getRandListOfSize(100);
consume(StreamMemoryTest.getIndexOfNothingStreamImpl(randomNums, x -> x < 0));
// or ...
// consume(StreamMemoryTest.getIndexOfNothingManualImpl(randomNums, x -> x < 0));
if (randomNums == null) {
break;
}
}
System.out.print(StreamMemoryTest.result());
}
}
流实现：
为TLAB分配的内存为64.62 GB
Class Average Object Size(bytes) Total Object Size(bytes) TLABs Average TLAB Size(bytes) Total TLAB Size(bytes) Pressure(%)
java.lang.Object[] 415.974 6,226,712 14,969 2,999,696.432 44,902,455,888 64.711
java.util.stream.ReferencePipeline$2 64 131,264 2,051 2,902,510.795 5,953,049,640 8.579
java.util.stream.ReferencePipeline$Head 56 72,744 1,299 3,070,768.043 3,988,927,688 5.749
java.util.stream.ReferencePipeline$2$1 24 25,128 1,047 3,195,726.449 3,345,925,592 4.822
java.util.Random 32 30,976 968 3,041,212.372 2,943,893,576 4.243
java.util.ArrayList 24 24,576 1,024 2,720,615.594 2,785,910,368 4.015
java.util.stream.FindOps$FindSink$OfRef 24 18,864 786 3,369,412.295 2,648,358,064 3.817
java.util.ArrayList$ArrayListSpliterator 32 14,720 460 3,080,696.209 1,417,120,256 2.042
手动实现：
为TLAB分配的内存46.06 GB
Class Average Object Size(bytes) Total Object Size(bytes) TLABs Average TLAB Size(bytes) Total TLAB Size(bytes) Pressure(%)
java.lang.Object[] 415.961 4,190,392 10,074 4,042,267.769 40,721,805,504 82.33
java.util.Random 32 32,064 1,002 4,367,131.521 4,375,865,784 8.847
java.util.ArrayList 24 14,976 624 3,530,601.038 2,203,095,048 4.454
还有其他人遇到过流对象本身消耗内存的问题吗？/这是已知的问题吗？
getIndexOfNothingManualImpl的等价实现可以是return nums.stream().anyMatch(predicate)吗？ - Zirconfor循环在幕后创建了一个Iterator实现。不知何故，您的分析工具错过了这一点... - HolgerIterator的实现）运行时，应该创建了更多的Random实例，但却创建了显著较少的ArrayList。你怎么能相信这样的数字呢？ - Holger