AggregateEventHandler.java
对EventHandler列表的封装,类似EventHandler List的功能,还实现了生命周期的管理,onStart onShutdown。
Sequence.java
Cache line padded sequence counter 补齐Cache line的序列计数器,ringbuffer和BatchEventProcessor使用到此类来计数。
补齐方式:
publiclong p1, p2, p3, p4, p5, p6, p7;// cache line padding, padding1 privatevolatilelong cursor = INITIAL_CURSOR_VALUE; publiclong p8, p9, p10, p11, p12, p13, p14;// cache line padding. padding2
情形1: object(0~8byte)+ padding1,
cursor+padding2
情形2: padding1+ cursor,
padding2 + object
这样,保证不同Sequence instance在不同的Cache line
参考资料:http://mechanical-sympathy.blogspot.com/2011/07/false-sharing.html
因为高速缓存是64字节,而Hotspot JVM的对象头是两个部分组成,第一部分是由24字节的hash code和8字节的锁等状态标识组成,第二部分是指向该对象类的引用。数组Array有一个附加的"word"来记录数组长度。每个对象为了性能优化,采用8个byte粒度边界对齐的。为了在packing的时候更高效,对象的field被从定义顺序(基于字节大小)按下列顺序重排:
1.doubles(8) and longs(8)
2.ints(4) and floats(4)
3.shorts(2) and chars(2)
4.booleans(1) and bytes(1)
5.references(4/8)
6.<repeat for sub-class fields>
所以我们补齐cache line:在任意field之间补上7个long(8)
BatchEventProcessor.java 批量从RingBuffer获取event代理給EventHandler处理。
关键代码:
public void run()
{
if (!running.compareAndSet(false, true))
{
throw new IllegalStateException("Thread is already running");
}
sequenceBarrier.clearAlert();
notifyStart();
T event = null;
long nextSequence = sequence.get() + 1L;
while (true)
{
try
{
final long availableSequence = sequenceBarrier.waitFor(nextSequence);
//批量处理,nextSequence无限增长怎么办?
while (nextSequence <= availableSequence)
{
event = ringBuffer.get(nextSequence);
eventHandler.onEvent(event, nextSequence,nextSequence == availableSequence);
nextSequence++;
}
sequence.set(nextSequence - 1L);//注意回退1,标示(nextSequence - 1L)的event已经消费完成
}
catch (final AlertException ex)
{
if (!running.get())
{
break;
}
}
catch (final Throwable ex)
{
exceptionHandler.handleEventException(ex,nextSequence, event);//异常处理类处理异常信息
sequence.set(nextSequence);//跳过异常信息的序列
nextSequence++;
}
}
notifyShutdown();
running.set(false);
}
ClaimStrategy.java
Sequencer里面的、用于event publishers申请event序列的策略合同。
有以下3种实现:
SingleThreadedClaimStrategy.java: 针对发布者的策略的单线程实现,只能在单线程做publisher的场景使用。
关键方法:
// availableCapacity 需要申请的可用数量
// dependentSequences 依赖的序列
public boolean hasAvailableCapacity(final int availableCapacity, final Sequence[] dependentSequences)
{
final long wrapPoint = (claimSequence.get() + availableCapacity) - bufferSize;//当前已经作为发布使用的序列(未被消费)+申请数量-
if (wrapPoint > minGatingSequence.get())
{
long minSequence = getMinimumSequence(dependentSequences);
//取出依赖序列中的最小的序列(未被消费)
minGatingSequence.set(minSequence);
if (wrapPoint > minSequence)
{
return false;//如果期望的到达的序列位置大于依赖序列中的最小的序列(未被消费),说明尚未消费,所以没有可用序列用于给发布者分配
}
}
return true;
}
private void waitForFreeSlotAt(final long sequence, final Sequence[] dependentSequences)
{
final long wrapPoint = sequence - bufferSize;
if (wrapPoint > minGatingSequence.get())
{
long minSequence;
while (wrapPoint > (minSequence = getMinimumSequence(dependentSequences)))
{
LockSupport.parkNanos(1L);//等待1纳秒
}
minGatingSequence.set(minSequence);
}
}
MultiThreadedClaimStrategy.java
@Override
public long incrementAndGet(final Sequence[] dependentSequences)
{
final MutableLong minGatingSequence = minGatingSequenceThreadLocal.get();
waitForCapacity(dependentSequences,minGatingSequence);//什么技巧?
final long nextSequence = claimSequence.incrementAndGet();
waitForFreeSlotAt(nextSequence,dependentSequences, minGatingSequence);
return nextSequence;
}
@Override
public long incrementAndGet(final int delta, final Sequence[] dependentSequences)
{
final long nextSequence = claimSequence.addAndGet(delta);
waitForFreeSlotAt(nextSequence,dependentSequences, minGatingSequenceThreadLocal.get());
return nextSequence;
}
@Override
public void serialisePublishing(final long sequence, final Sequence cursor, final int batchSize)
{
int counter = RETRIES;
while (sequence - cursor.get() > pendingPublication.length())
{
if (--counter == 0)
{
Thread.yield();
counter = RETRIES;
}
}
long expectedSequence = sequence - batchSize;
for (long pendingSequence = expectedSequence + 1;pendingSequence <= sequence; pendingSequence++)
{
pendingPublication.set((int) pendingSequence& pendingMask, pendingSequence);
}
long cursorSequence = cursor.get();
if (cursorSequence >= sequence)
{
return;
}
expectedSequence = Math.max(expectedSequence,cursorSequence);
long nextSequence = expectedSequence + 1;
while (cursor.compareAndSet(expectedSequence, nextSequence))
{
expectedSequence = nextSequence;
nextSequence++;
if (pendingPublication.get((int) nextSequence & pendingMask) != nextSequence) //这里是什么含义?只有当nextSequence 大于 PendingBufferSize才会出现不相等的情况。
{
break;
}
}
}
MultiThreadedLowContentionClaimStrategy.java
与MultiThreadedClaimStrategy.java的在于:
@Override
public void serialisePublishing(final long sequence, final Sequence cursor, final int batchSize)
{
final long expectedSequence = sequence - batchSize;
while (expectedSequence != cursor.get())//会不会死循环?
{
// busy spin
}
cursor.set(sequence);
}
EventPublisher.java
时间发布者,主要代码:
private void translateAndPublish(final EventTranslator<E> translator, final long sequence)
{
try
{
translator.translateTo(ringBuffer.get(sequence),sequence);//需要根据传入的translator来依据sequence转换event后,再发布event.
}
finally
{
ringBuffer.publish(sequence);
}
}
WaitStrategy.java
定制EventProcessor等待cursor这个sequence的策略,有以下4种实现:
/**
* Blocking strategy that uses a lock andcondition variable for {@linkEventProcessor}s waiting on a barrier.
*
* This strategy can be used when throughputand low-latencyare not as important as CPU resource.
*/
BlockingWaitStrategy.java:用到了lock,所以只适合用在throughput和low-latency要求不高的情况下。
/**
* Busy Spin strategy that uses a busy spinloop for {@link com.lmax.disruptor.EventProcessor}s waiting on a barrier.
*
* This strategy will use CPU resource to avoidsyscalls which can introduce latency jitter. It is best
* used when threads can be bound to specificCPU cores.
*/
BusySpinWaitStrategy.java:这种是耗cpu的做法,不做yield()。
/**
* Sleeping strategy that initially spins, thenuses a Thread.yield(), and eventually for the minimum number of nanos
* the OS and JVM will allow while the {@link com.lmax.disruptor.EventProcessor}s are waiting on a barrier.
*
* This strategy is a good compromise betweenperformance and CPU resource. Latency spikes can occur after quiet periods.
*/
SleepingWaitStrategy.java:做一个counter的判断,小于100才yield(),小于0做LockSupport.parkNanos(1L);
/**
* Yielding strategy that uses a Thread.yield()for {@link com.lmax.disruptor.EventProcessor}s waiting on a barrier
* after an initially spinning.
*
* This strategy is a good compromise betweenperformance and CPU resource without incurring significant latency spikes.
*/
YieldingWaitStrategy.java:counter==0,才做yield()。