高并发之限流RateLimiter（二）

Guava RateLimiter提供了令牌桶算法实现：平滑突发限流(SmoothBursty)和平滑预热限流(SmoothWarmingUp)实现。

SmoothBursty：令牌生成速度恒定

 @Test
    public void testAcquire() {
        // acquire(i); 获取令牌，返回阻塞的时间,支持预消费.
        RateLimiter limiter = RateLimiter.create(1);

        for (int i = 1; i < 10; i++) {
            double waitTime = limiter.acquire();
            System.out.println("cutTime=" + longToDate(System.currentTimeMillis()) + " acq:" + i + " waitTime:" + waitTime);
        }
    }

    public static String longToDate(long lo){
        Date date = new Date(lo);
        SimpleDateFormat sd = new SimpleDateFormat("yyyy-MM-dd HH:mm:ss");
        return sd.format(date);
    }

输出结果：

cutTime=2019-03-29 09:31:42 acq:1 waitTime:0.0
cutTime=2019-03-29 09:31:43 acq:2 waitTime:0.989135
cutTime=2019-03-29 09:31:44 acq:3 waitTime:0.998023
cutTime=2019-03-29 09:31:45 acq:4 waitTime:0.999573
cutTime=2019-03-29 09:31:46 acq:5 waitTime:0.999359
cutTime=2019-03-29 09:31:47 acq:6 waitTime:0.999566
cutTime=2019-03-29 09:31:48 acq:7 waitTime:0.998763
cutTime=2019-03-29 09:31:49 acq:8 waitTime:0.999163
cutTime=2019-03-29 09:31:50 acq:9 waitTime:1.000036

说明：每秒1个令牌生成一个令牌，从输出可看出很平滑，这种实现将突发请求速率平均成固定请求速率。

下面demo是突发请求：

@Test
    public void testAcquire2() {
        // 请求突发
        RateLimiter limiter = RateLimiter.create(5);

        for (int i = 1; i < 5; i++) {
            double waitTime = 0;
            if(i == 2){
                waitTime = limiter.acquire(10);
            }else{
                waitTime = limiter.acquire(1);
            }

            System.out.println("cutTime=" + longToDate(System.currentTimeMillis()) + " acq:" + i + " waitTime:" + waitTime);
        }
    }

输出：

cutTime=2019-03-29 09:53:55 acq:1 waitTime:0.0
cutTime=2019-03-29 09:53:56 acq:2 waitTime:0.188901
cutTime=2019-03-29 09:53:58 acq:3 waitTime:1.99789
cutTime=2019-03-29 09:53:58 acq:4 waitTime:0.198832

说明：

i=1,消费i个令牌，此时还剩4个令牌；

i=2，突发10个请求，令牌桶算法也允许了这种突发（允许消费未来的令牌）；

i=3，上次请求消费了，所以需要等待2s；

下面看源码：

---

简单介绍下：Stopwatch

public final class Stopwatch {
    private final Ticker ticker;//计时器，用于获取当前时间
    private boolean isRunning;//计时器是否运行中的状态标记
    private long elapsedNanos;//用于标记从计时器开启到调用统计的方法时过去的时间
    private long startTick;//计时器开启的时刻时间

    private long elapsedNanos() {
        return this.isRunning ? this.ticker.read() - this.startTick + this.elapsedNanos : this.elapsedNanos;
    }
    public long elapsed(TimeUnit desiredUnit) {
        return desiredUnit.convert(this.elapsedNanos(), TimeUnit.NANOSECONDS);
    }
 }

TimeUnit:

MILLISECONDS {
        public long toNanos(long d)   { return x(d, C2/C0, MAX/(C2/C0)); }
        public long toMicros(long d)  { return x(d, C2/C1, MAX/(C2/C1)); }
        public long toMillis(long d)  { return d; }
        public long toSeconds(long d) { return d/(C3/C2); }
        public long toMinutes(long d) { return d/(C4/C2); }
        public long toHours(long d)   { return d/(C5/C2); }
        public long toDays(long d)    { return d/(C6/C2); }
        public long convert(long d, TimeUnit u) { return u.toMillis(d); }
        int excessNanos(long d, long m) { return 0; }
    },
 
 MICROSECONDS {
        public long toNanos(long d)   { return x(d, C1/C0, MAX/(C1/C0)); }
        public long toMicros(long d)  { return d; }
        public long toMillis(long d)  { return d/(C2/C1); }
        public long toSeconds(long d) { return d/(C3/C1); }
        public long toMinutes(long d) { return d/(C4/C1); }
        public long toHours(long d)   { return d/(C5/C1); }
        public long toDays(long d)    { return d/(C6/C1); }
        public long convert(long d, TimeUnit u) { return u.toMicros(d); }
        int excessNanos(long d, long m) { return (int)((d*C1) - (m*C2)); }
    },
    
NANOSECONDS {
        public long toNanos(long d)   { return d; }
        public long toMicros(long d)  { return d/(C1/C0); }
        public long toMillis(long d)  { return d/(C2/C0); }
        public long toSeconds(long d) { return d/(C3/C0); }
        public long toMinutes(long d) { return d/(C4/C0); }
        public long toHours(long d)   { return d/(C5/C0); }
        public long toDays(long d)    { return d/(C6/C0); }
        public long convert(long d, TimeUnit u) { return u.toNanos(d); }
        int excessNanos(long d, long m) { return (int)(d - (m*C2)); }
},

其中：

static final long C0 = 1L;
static final long C1 = C0 * 1000L;
static final long C2 = C1 * 1000L;
static final long C3 = C2 * 1000L;
static final long C4 = C3 * 60L;
static final long C5 = C4 * 60L;
static final long C6 = C5 * 24L;

@Test
public void stopwatch1() {
    Stopwatch stopwatch = Stopwatch.createStarted();

    doSomething();
    stopwatch.stop(); // optional
    long millis = stopwatch.elapsed(MILLISECONDS);
    System.out.println("time: " + stopwatch);
}

@Test
public void stopwatch2() {
    Stopwatch stopwatch = Stopwatch.createStarted();
    //doSomething();
    stopwatch.stop();
    long millis = stopwatch.elapsed(MILLISECONDS);
    System.out.println("time: " + stopwatch);

    stopwatch.reset().start();
    //doSomething();
    stopwatch.stop();
    millis = stopwatch.elapsed(MILLISECONDS);
    System.out.println("time: " + stopwatch);
}

public static void doSomething(){
    try {
        Thread.sleep(100);
    } catch (InterruptedException e) {
        e.printStackTrace();
    }
}

stopwatch1结果：

time: 100.8 ms

执行过程：

使用stopwatch对程序运行时间进行调试，首先调用StopWatch.createStarted()创建并启动一个stopwatch实例，调用stopwatch.stop()停止计时，此时会更新stopwatch的elapsedNanos时间，为stopwatch开始启动到结束计时的时间，再次调用stopwatch.elapsed()，获取stopwatch在start-stop时间段，时间流逝的长度。

---

RateLimiter.class

public static RateLimiter create(double permitsPerSecond) {
        return create(permitsPerSecond, RateLimiter.SleepingStopwatch.createFromSystemTimer());//Stopwatch类稍后
    }

    @VisibleForTesting
    static RateLimiter create(double permitsPerSecond, RateLimiter.SleepingStopwatch stopwatch) {
        RateLimiter rateLimiter = new SmoothBursty(stopwatch, 1.0D);
        rateLimiter.setRate(permitsPerSecond);
        return rateLimiter;
    }
    
    public final void setRate(double permitsPerSecond) {
        Preconditions.checkArgument(permitsPerSecond > 0.0D && !Double.isNaN(permitsPerSecond), "rate must be positive");
        synchronized(this.mutex()) {
            this.doSetRate(permitsPerSecond, this.stopwatch.readMicros());
        }
    }
    
    abstract void doSetRate(double var1, long var3);

说明：this.stopwatch.readMicros());源码最终调用的是

NANOSECONDS {
    public long toNanos(long d)   { return d; }
    public long toMicros(long d)  { return d/(C1/C0); } //return (stopwatch中的elapsedNanos，表示时间差)/(1L * 1000L/1L)    
}

SmoothRateLimiter

final void doSetRate(double permitsPerSecond, long nowMicros) {
    this.resync(nowMicros);
    double stableIntervalMicros = (double)TimeUnit.SECONDS.toMicros(1L) / permitsPerSecond;
    this.stableIntervalMicros = stableIntervalMicros;
    this.doSetRate(permitsPerSecond, stableIntervalMicros);
}
abstract void doSetRate(double var1, double var3);

void resync(long nowMicros) {
    if (nowMicros > this.nextFreeTicketMicros) {
        //相当于(double)(nowMicros - this.nextFreeTicketMicros) * (permitsPerSecond double)TimeUnit.SECONDS.toMicros(1L)) //令牌生成速率：xx/单位时间
        double newPermits = (double)(nowMicros - this.nextFreeTicketMicros) / this.coolDownIntervalMicros();
        this.storedPermits = Math.min(this.maxPermits, this.storedPermits + newPermits);
        this.nextFreeTicketMicros = nowMicros;
    }
}

说明：

nowMicros：表示用于标记从计时器开启到调用统计的方法时过去的时间
coolDownIntervalMicros：添加令牌时间间隔
stableIntervalMicros：添加令牌时间间隔 = (double)TimeUnit.SECONDS.toMicros(1L) / permitsPerSecond；(1秒/每秒的令牌数)
newPermits：时间段内新生令牌数
storedPermits：当前令牌数

nextFreeTicketMicros：

下一次请求可以获取令牌的起始时间，由于RateLimiter允许预消费，上次请求预消费令牌后，下次请求需要等待相应的时间到nextFreeTicketMicros时刻才可以获取令牌

SmoothBursty

static final class SmoothBursty extends SmoothRateLimiter {
        final double maxBurstSeconds;

        SmoothBursty(SleepingStopwatch stopwatch, double maxBurstSeconds) {
            super(stopwatch, null);
            this.maxBurstSeconds = maxBurstSeconds;//在RateLimiter未使用时，最多存储几秒的令牌
        }

        void doSetRate(double permitsPerSecond, double stableIntervalMicros) {
            double oldMaxPermits = this.maxPermits;
            this.maxPermits = this.maxBurstSeconds * permitsPerSecond;
            if (oldMaxPermits == 1.0D / 0.0) { //相当于oldMaxPermits ==Double.POSITIVE_INFINITY ，Double.POSITIVE_INFINITY 表示无穷大
                
                this.storedPermits = this.maxPermits;
            } else {
                this.storedPermits = oldMaxPermits == 0.0D ? 0.0D : this.storedPermits * this.maxPermits / oldMaxPermits;
            }

        }

        long storedPermitsToWaitTime(double storedPermits, double permitsToTake) {
            return 0L;
        }

        double coolDownIntervalMicros() {
            return this.stableIntervalMicros;
        }
    }

参数说明：

maxBurstSeconds:在RateLimiter未使用时，最多存储几秒的令牌
permitsPerSecond: 速率=令牌数/每秒
maxPermits :最大存储令牌数 = maxBurstSeconds * permitsPerSecond
storedPermits： 当前存储令牌数

RateLimiter几个常用接口分析

1、acquire() 函数主要用于获取permits个令牌，并计算需要等待多长时间，进而挂起等待，并将该值返回

RateLimiter.calss

@CanIgnoreReturnValue
public double acquire() {
  return acquire(1);
}

/**
* 获取令牌，返回阻塞的时间
**/
@CanIgnoreReturnValue
public double acquire(int permits) {
  long microsToWait = reserve(permits);
  
  //获取等待时间后，阻塞线程
  stopwatch.sleepMicrosUninterruptibly(microsToWait);
  return 1.0 * microsToWait / SECONDS.toMicros(1L);
}

final long reserve(int permits) {
  checkPermits(permits);
  synchronized (mutex()) {
    return reserveAndGetWaitLength(permits, stopwatch.readMicros());
  }
}

final long reserveAndGetWaitLength(int permits, long nowMicros) {
    long momentAvailable = this.reserveEarliestAvailable(permits, nowMicros);
    return Math.max(momentAvailable - nowMicros, 0L);
}
abstract long reserveEarliestAvailable(int var1, long var2);

SmoothRateLimiter.class

final long reserveEarliestAvailable(int requiredPermits, long nowMicros) {
        this.resync(nowMicros);
        long returnValue = this.nextFreeTicketMicros;//resync()方法后，如果nowMicros > this.nextFreeTicketMicros,等于nowMicros
        
        double storedPermitsToSpend = Math.min((double)requiredPermits, this.storedPermits);
        //freshPermits从令牌桶中获取令牌后还需要的令牌数量
        double freshPermits = (double)requiredPermits - storedPermitsToSpend;
        
        //平滑这里this.storedPermitsToWaitTime()直接返回0L + 还需要令牌数量/速率（需要的时间）
        long waitMicros = this.storedPermitsToWaitTime(this.storedPermits, storedPermitsToSpend) + (long)(freshPermits * this.stableIntervalMicros);
        
        //如果超前消费，将导致下次请求等待时间=LongMath.saturatedAdd(this.nextFreeTicketMicros, waitMicros);
        this.nextFreeTicketMicros = LongMath.saturatedAdd(this.nextFreeTicketMicros, waitMicros);
        this.storedPermits -= storedPermitsToSpend;
        return returnValue;
    }

2、tryAcquire()

函数可以尝试在timeout时间内获取令牌，如果可以则挂起等待相应时间并返回true，否则立即返回false

 public boolean tryAcquire(int permits, long timeout, TimeUnit unit) {
        long timeoutMicros = Math.max(unit.toMicros(timeout), 0L);//超时时间
        checkPermits(permits);
        long microsToWait;
        synchronized(this.mutex()) {
            long nowMicros = this.stopwatch.readMicros();
            if (!this.canAcquire(nowMicros, timeoutMicros)) {
                return false;
            }
            //获取需要阻塞时间
            microsToWait = this.reserveAndGetWaitLength(permits, nowMicros);
        }

        this.stopwatch.sleepMicrosUninterruptibly(microsToWait);
        return true;
    }

    private boolean canAcquire(long nowMicros, long timeoutMicros) {
        //下一次请求可以获取令牌的起始时间
        return this.queryEarliestAvailable(nowMicros) - timeoutMicros <= nowMicros;
    }

canAcquire用于判断timeout时间内是否可以获取令牌，通过判断当前时间+超时时间是否大于nextFreeTicketMicros 来决定是否能够拿到足够的令牌数，如果可以获取到，则过程同acquire，线程sleep等待，如果通过canAcquire在此超时时间内不能回去到令牌，则可以快速返回，不需要等待timeout后才知道能否获取到令牌。

SmoothWarmingUp:令牌生成速度缓慢提升直到维持在一个稳定值

SmoothWarmingUp创建方式：RateLimiter.create(doublepermitsPerSecond, long warmupPeriod, TimeUnit unit)

permitsPerSecond表示每秒新增的令牌数，warmupPeriod表示在从冷启动速率过渡到平均速率的时间间隔。

@Test
    public void acquire1() {
        RateLimiter limiter = RateLimiter.create(5, 1000, TimeUnit.MILLISECONDS);
        for (int i = 1; i < 6; i++) {
            System.out.println(limiter.acquire());
        }

        try {
            Thread.sleep(1000L);
        } catch (InterruptedException e) {
            e.printStackTrace();
        }

        for (int i = 1; i < 6; i++) {
            System.out.println(limiter.acquire());
        }
    }

 结果：

0.0
0.518741
0.357811
0.219877
0.199584
0.0
0.361189
0.220761
0.19938
0.199856

速率是梯形上升速率的，也就是说冷启动时会以一个比较大的速率慢慢到平均速率；然后趋于平均速率（梯形下降到平均速率）。可以通过调节warmupPeriod参数实现一开始就是平滑固定速率。

参考：

https://www.cnblogs.com/xuwc/p/9123078.html

https://www.cnblogs.com/xuwc/p/9123078.html