Java无锁性能JMH

Question

我有一个JMH多线程测试:

@State(Scope.Benchmark)
@BenchmarkMode(Mode.Throughput)
@OutputTimeUnit(TimeUnit.MICROSECONDS)
@Fork(value = 1, jvmArgsAppend = { "-Xmx512m", "-server", "-XX:+AggressiveOpts","-XX:+UnlockDiagnosticVMOptions",
        "-XX:+UnlockExperimentalVMOptions", "-XX:+PrintAssembly", "-XX:PrintAssemblyOptions=intel",
        "-XX:+PrintSignatureHandlers"})
@Measurement(iterations = 5, time = 5, timeUnit = TimeUnit.SECONDS)
@Warmup(iterations = 3, time = 2, timeUnit = TimeUnit.SECONDS)
public class LinkedQueueBenchmark {
private static final Unsafe unsafe = UnsafeProvider.getUnsafe();
private static final long offsetObject;
private static final long offsetNext;

private static final int THREADS = 5;
private static class Node {
    private volatile Node next;
    public Node() {}
}

static {
    try {
        offsetObject = unsafe.objectFieldOffset(LinkedQueueBenchmark.class.getDeclaredField("object"));
        offsetNext = unsafe.objectFieldOffset(Node.class.getDeclaredField("next"));
    } catch (Exception ex) { throw new Error(ex); }
}

protected long t0,t1,t2,t3,t4,t5,t6,t7;
private volatile Node object = new Node(null);


@Threads(THREADS)
@Benchmark
public Node doTestCasSmart() {
    Node current, o = new Node();
    for(;;) {
        current = this.object;
        if (unsafe.compareAndSwapObject(this, offsetObject, current, o)) {
            //current.next = o; //Special line:
            break;
        } else {
            LockSupport.parkNanos(1);
        }
    }
    return current;
}
}

1. 在目前的变体中,我的表现为~55 ops/us
2. 但是,如果我取消注释"特殊线",或用unsafe.putOrderedObject(在任何方向 - current.next = o或o.next = current)替换它,性能~2 ops/us.

据我所知,这是CPU缓存发生的事情,也许它正在清理存储缓冲区.如果我将它替换为基于锁的方法,没有CAS,性能将是11-20 ops/us.
我尝试使用LinuxPerfAsmProfiler和PrintAssembly,在第二种情况下,我看到:

....[Hottest Regions]...............................................................................
 25.92%   17.93%  [0x7f1d5105fe60:0x7f1d5105fe69] in SpinPause (libjvm.so)
 17.53%   20.62%  [0x7f1d5119dd88:0x7f1d5119de57] in ParMarkBitMap::live_words_in_range(HeapWord*, oopDesc*) const (libjvm.so)
 10.81%    6.30%  [0x7f1d5129cff5:0x7f1d5129d0ed] in ParallelTaskTerminator::offer_termination(TerminatorTerminator*) (libjvm.so)
  7.99%    9.86%  [0x7f1d3c51d280:0x7f1d3c51d3a2] in com.jad.generated.LinkedQueueBenchmark_doTestCasSmart::doTestCasSmart_thrpt_jmhStub 

有人可以向我解释一下究竟发生了什么？为什么这么慢？这里存储装载障碍？为什么putOrdered不起作用？以及如何解决它？

Answer 1

规则:你应该首先寻找愚蠢的错误,而不是寻找"高级"答案.

SpinPause,ParMarkBitMap::live_words_in_range(HeapWord*, oopDesc*)以及ParallelTaskTerminator::offer_termination(TerminatorTerminator*)来自GC线程.这很可能意味着大多数工作基准测试都是GC.实际上,运行"特殊线"没有注释-prof gc产量:

# Run complete. Total time: 00:00:43

Benchmark                      Mode  Cnt      Score    Error   Units
LQB.doTestCasSmart            thrpt    5      5.930 ±  3.867  ops/us
LQB.doTestCasSmart:·gc.time   thrpt    5  29970.000               ms

因此,在43秒的运行中,你花了30秒做GC.或者,甚至平原-verbose:gc显示它:

Iteration   3: [Full GC (Ergonomics)  408188K->1542K(454656K), 0.0043022 secs]
[GC (Allocation Failure)  60422K->60174K(454656K), 0.2061024 secs]
[GC (Allocation Failure)  119054K->118830K(454656K), 0.2314572 secs]
[GC (Allocation Failure)  177710K->177430K(454656K), 0.2268396 secs]
[GC (Allocation Failure)  236310K->236054K(454656K), 0.1718049 secs]
[GC (Allocation Failure)  294934K->294566K(454656K), 0.2265855 secs]
[Full GC (Ergonomics)  294566K->147408K(466432K), 0.7139546 secs]
[GC (Allocation Failure)  206288K->205880K(466432K), 0.2065388 secs]
[GC (Allocation Failure)  264760K->264312K(466432K), 0.2314117 secs]
[GC (Allocation Failure)  323192K->323016K(466432K), 0.2183271 secs]
[Full GC (Ergonomics)  323016K->322663K(466432K), 2.8058725 secs]

2.8s完整的GC,这很糟糕.大约5s花在GC上,在一个由5s运行时间限制的迭代中.那太糟糕了.

为什么？好吧,你正在那里建立链表.当然,队列的头部是无法到达的,object应该收集从头到头的所有内容.但收藏不是即时的.队列越长,消耗的内存越多,GC遍历它的工作量就越大.这是一个积极的反馈循环,削弱了执行力.由于队列元素无论如何都是可收集的,因此这个反馈循环永远不会到达OOME.object在新head领域中存储初始内容将最终对OOME进行测试.

因此,putOrdered坦率地说,您的问题与内存障碍或队列性能无关.我想你需要重新考虑你实际测试的是什么.设计测试使每次@Benchmark调用瞬态内存占用保持不变是一门艺术.