为什么“-XX:+UseLWPSynchronization”会对 Windows OS 上的 stamplock 产生负面影响？

Question

我想在 java 中实现一个简单的速率限制器来学习如何使用 jmh。在 'https://github.com/William1104/rate-limiter'

处创建了一个简单的 github 项目

有趣的是，当启用“-XX:+UseLWPSynchronization”选项时，某些实现（使用 stamplock）的吞吐量会受到影响。基准测试是在 Windows 机器上进行的，我希望它对非 Solaris 系统没有影响。但是，测试结果显示不同。我可以知道有人可以帮助我了解到底发生了什么吗？

以下是我机器上的测试结果作为参考：

有选项： -server, -XX:+UnlockDiagnosticVMOptions, -XX:+UseNUMA

Benchmark	(rateLimiterType)	Mode	Cnt	Score	Error	Units
RaterLimiterBenchmark.thread_1	StampLockLongArrayRateLimiter	thrpt	90	21487.385	▒ 1082.163	ops/ms
RaterLimiterBenchmark.thread_1	StampLockInstantArrayRateLimiter	thrpt	90	13162.330	▒ 1585.555	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedLongArrayRateLimiter	thrpt	90	15362.934	▒ 227.704	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedInstantArrayRateLimiter	thrpt	90	17281.675	▒ 2148.057	ops/ms
RaterLimiterBenchmark.thread_10	StampLockLongArrayRateLimiter	thrpt	90	6868.653	▒ 146.372	ops/ms
RaterLimiterBenchmark.thread_10	StampLockInstantArrayRateLimiter	thrpt	90	8189.747	▒ 335.517	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedLongArrayRateLimiter	thrpt	90	6643.004	▒ 103.568	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedInstantArrayRateLimiter	thrpt	90	5252.975	▒ 190.363	ops/ms
RaterLimiterBenchmark.thread_100	StampLockLongArrayRateLimiter	thrpt	90	7352.890	▒ 2109.446	ops/ms
RaterLimiterBenchmark.thread_100	StampLockInstantArrayRateLimiter	thrpt	90	8675.814	▒ 922.653	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedLongArrayRateLimiter	thrpt	90	6509.368	▒ 157.212	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedInstantArrayRateLimiter	thrpt	90	5042.867	▒ 192.971	ops/ms

有选项：-server, -XX:+UnlockDiagnosticVMOptions, -XX:+UseNUMA, -XX:+UseLWPSynchronization

Benchmark	(rateLimiterType)	Mode	Cnt	Score	Error	Units
RaterLimiterBenchmark.thread_1	StampLockLongArrayRateLimiter	thrpt	90	11383.198	▒ 353.921	ops/ms
RaterLimiterBenchmark.thread_1	StampLockInstantArrayRateLimiter	thrpt	90	11666.918	▒ 842.426	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedLongArrayRateLimiter	thrpt	90	15696.852	▒ 371.078	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedInstantArrayRateLimiter	thrpt	90	15357.617	▒ 650.846	ops/ms
RaterLimiterBenchmark.thread_10	StampLockLongArrayRateLimiter	thrpt	90	6937.050	▒ 130.727	ops/ms
RaterLimiterBenchmark.thread_10	StampLockInstantArrayRateLimiter	thrpt	90	8268.909	▒ 291.471	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedLongArrayRateLimiter	thrpt	90	9134.319	▒ 1208.998	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedInstantArrayRateLimiter	thrpt	90	5294.341	▒ 225.995	ops/ms
RaterLimiterBenchmark.thread_100	StampLockLongArrayRateLimiter	thrpt	90	8453.825	▒ 1075.312	ops/ms
RaterLimiterBenchmark.thread_100	StampLockInstantArrayRateLimiter	thrpt	90	16297.921	▒ 611.255	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedLongArrayRateLimiter	thrpt	90	12536.378	▒ 974.951	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedInstantArrayRateLimiter	thrpt	90	9051.560	▒ 1303.856	ops/ms

有StampLockLongArrayRateLimiter和SynchronizedLongArrayRateLImiter的实现：

package one.williamwong.ratelimiter;

import java.time.Duration;
import java.util.Arrays;
import java.util.concurrent.locks.StampedLock;

public class StampLockLongArrayRateLimiter implements IRateLimiter {

    private final long duration;
    private final long[] records;
    private final StampedLock lock;
    private int pointer;

    public StampLockLongArrayRateLimiter(int maxInvokes, Duration duration) {
        this.duration = duration.toNanos();
        this.records = new long[maxInvokes];
        this.lock = new StampedLock();
        this.pointer = 0;
    }

    @Override public void acquire() {
        final long stamp = lock.writeLock();
        try {
            final long now = System.nanoTime();
            if (records[pointer] != 0) {
                final long awayFromHead = now - records[pointer];
                if (awayFromHead < duration) {
                    handleExcessLimit(records.length, Duration.ofNanos(awayFromHead));
                }
            }
            records[pointer] = now;
            pointer = (pointer + 1) % records.length;
        } finally {
            lock.unlockWrite(stamp);
        }
    }

    @Override public void reset() {
        final long stamp = lock.writeLock();
        try {
            Arrays.fill(records, 0);
            this.pointer = 0;
        } finally {
            lock.unlockWrite(stamp);
        }
    }

}

package one.williamwong.ratelimiter;

import java.time.Duration;
import java.util.Arrays;

public class SynchronizedLongArrayRateLimiter implements IRateLimiter {

    private final long duration;
    private final long[] records;
    private final Object lock;
    private int pointer;

    public SynchronizedLongArrayRateLimiter(int maxInvokes, Duration duration) {
        this.duration = duration.toNanos();
        this.records = new long[maxInvokes];
        this.lock = new Object();
        this.pointer = 0;
    }

    @Override
    public void acquire() {
        synchronized (lock) {
            final long now = System.nanoTime();
            if (records[pointer] != 0) {
                final long awayFromHead = now - records[pointer];
                if (awayFromHead < duration) {
                    handleExcessLimit(records.length, Duration.ofNanos(awayFromHead));
                }
            }
            records[pointer] = now;
            pointer = (pointer + 1) % records.length;
        }
    }

    @Override public void reset() {
        synchronized (lock) {
            Arrays.fill(records, 0);
            this.pointer = 0;
        }
    }

}

Answer 1

感谢您的评论。我用不同的设置重新运行我的基准测试。如果我们每次迭代 1 秒再次执行 JMH，我得到以下结果：

带有选项：-server、-XX:+UnlockDiagnosticVMOptions、-XX:+UseNUMA、-XX:-UseLWPSynchronization

Benchmark	(rateLimiterType)	Mode	Cnt	Score	Error	Units
RaterLimiterBenchmark.thread_1	StampLockLongArrayRateLimiter	thrpt	90	23573.282	▒ 364.739	ops/ms
RaterLimiterBenchmark.thread_1	StampLockInstantArrayRateLimiter	thrpt	90	23062.260	▒ 1035.395	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedLongArrayRateLimiter	thrpt	90	34667.411	▒ 246.003	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedInstantArrayRateLimiter	thrpt	90	36426.369	▒ 1248.360	ops/ms
RaterLimiterBenchmark.thread_10	StampLockLongArrayRateLimiter	thrpt	90	13592.158	▒ 76.319	ops/ms
RaterLimiterBenchmark.thread_10	StampLockInstantArrayRateLimiter	thrpt	90	14564.306	▒ 474.613	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedLongArrayRateLimiter	thrpt	90	13524.610	▒ 155.850	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedInstantArrayRateLimiter	thrpt	90	13080.967	▒ 309.736	ops/ms
RaterLimiterBenchmark.thread_100	StampLockLongArrayRateLimiter	thrpt	90	13224.529	▒ 459.035	ops/ms
RaterLimiterBenchmark.thread_100	StampLockInstantArrayRateLimiter	thrpt	90	13890.278	▒ 456.182	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedLongArrayRateLimiter	thrpt	90	12672.925	▒ 314.118	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedInstantArrayRateLimiter	thrpt	90	12245.120	▒ 296.395	ops/ms

有选项：-server、-XX:+UnlockDiagnosticVMOptions、-XX:+UseNUMA、-XX:+UseLWPSynchronization

Benchmark	(rateLimiterType)	Mode	Cnt	Score	Error	Units
RaterLimiterBenchmark.thread_1	StampLockLongArrayRateLimiter	thrpt	90	24842.514	▒ 372.521	ops/ms
RaterLimiterBenchmark.thread_1	StampLockInstantArrayRateLimiter	thrpt	90	24327.864	▒ 322.659	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedLongArrayRateLimiter	thrpt	90	34490.411	▒ 330.288	ops/ms
RaterLimiterBenchmark.thread_1	SynchronizedInstantArrayRateLimiter	thrpt	90	38383.257	▒ 654.269	ops/ms
RaterLimiterBenchmark.thread_10	StampLockLongArrayRateLimiter	thrpt	90	13536.284	▒ 74.613	ops/ms
RaterLimiterBenchmark.thread_10	StampLockInstantArrayRateLimiter	thrpt	90	13702.022	▒ 289.616	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedLongArrayRateLimiter	thrpt	90	12530.107	▒ 243.471	ops/ms
RaterLimiterBenchmark.thread_10	SynchronizedInstantArrayRateLimiter	thrpt	90	10795.833	▒ 158.400	ops/ms
RaterLimiterBenchmark.thread_100	StampLockLongArrayRateLimiter	thrpt	90	13204.275	▒ 200.937	ops/ms
RaterLimiterBenchmark.thread_100	StampLockInstantArrayRateLimiter	thrpt	90	11606.823	▒ 224.213	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedLongArrayRateLimiter	thrpt	90	11504.124	▒ 107.543	ops/ms
RaterLimiterBenchmark.thread_100	SynchronizedInstantArrayRateLimiter	thrpt	90	10732.451	▒ 118.753	ops/ms

无论是否启用 'UseLWPSynchronization'，我都没有观察到巨大的性能差异。我遇到的问题与不稳定的 JMH 设置有关。

为什么“-XX:+UseLWPSynchronization”会对 Windows OS 上的 stamplock 产生负面影响？

Why could '-XX:+UseLWPSynchronization' negatively impact stamplock on Windows OS?

java

synchronization

jvm

performance-testing

jmh