您现在的位置：首页 --> 系统架构 --> 消息分发的同步均衡策略

消息分发的同步均衡策略

浏览:5048次出处信息

TimeTunnel在做消息分发时有这样一个场景:

A类消息需要做实时分析, 且量很大, 故它的消费者不会只是一台机器, 而是一组机器, 并要求这组中每台机器收到的消息量应该平均的, 即A消息在某个时刻有100条, 若有4台机器消费的话, 最佳的情况每台机器应收到25条.

这个场景就好比, 一个消息队列, 有多个线程并行消费, 如何保证每个消费线程获取的消息数量一样的.

解决它的切入点可能有两个:

消费线程拉的方式, 这就需要一个协调器, 来同步每个消费线程的进度, 即每个消费线程每次拉完都需要询问协调器是否可以继续;

单一线程推的方式, 创建一组与消费线程数量一致的二级队列(即每个消费线程仅从与之绑定的队列中取); 独立一个分派线程, 它从一级队列里取消息, 然后轮转的推给每个二级消费队列.

这两条思路各有优缺点, 要视系统在时间和空间上的取舍而定. TimeTunnel 选择了第一种方式, 实现分了三步:

领域对象

消费进展(Track)

记录每个线程的当前消费状态.

同步误差范围(Range)

任意时刻可容忍的消费线程间消息数量的正负差距, 这个范围直接影响消费的效率, 范围越小则各个消费越均衡, 但同步牺牲了性能.

协调者(Coordinator)

根据每个消费线程的消费进展, 调控消费线程是活动还是等待.

01 /**
02  * {@link Coordinator} balance movings of {@link Track}s.
03  * 
04  * @author jushi
05  * @created 2010-12-14
06  * 
07  */
08 public interface Coordinator {
09   /**
10    * @param name
11    * @return a new {@link Track} with name.
12    */
13   Track track(String name);
14 
15   /**
16    * {@link Track} record movings.
17    */
18   public interface Track {
19     /**
20      * Destory {@link Track} from {@link Coordinator}.
21      */
22     void destory();
23 
24     /**
25      * Move one step.
26      * 
27      * @return false means current {@link Track} need wait for balance, else
28      *         true.
29      * @throws IllegalStateException if track destoryed.
30      */
31     boolean move();
32   }
33 }

验收用例

假定现在要协调两个人的步调, 每个人通过Track来汇报自己的走动的次数, 再根据汇报的反馈决定是暂定还是继续, 如下代码:

01   private final class Person implements Callable {
02     public Person(final Track track) {
03       this.track = track;
04     }
05 
06     @Override
07     public Integer call() throws Exception {
08       int steps = 0;
09       while (running.get()) {
10         if (canWalk()) steps++;
11       }
12       track.destory();
13       return steps;
14     }
15 
16     boolean canWalk() {
17       return track.move();
18     }
19 
20     private final Track track;
21   }

接下来, 就是要验证两人走了一段时间后, 二者步数的差距应该不超过预设的范围, 则有:

01   @Test
02   public void shouldSynchronizedTwoWalker() throws Exception {
03     final int value = 10;
04     final Coordinator coordinator = Coordinators.coordinator(value);
05 
06     final List> futures =
07       Race.run(new Person(coordinator.track("0")),
08                new Person(coordinator.track("1")),
09                new DelayHaltCaller());
10     final int steps0 = futures.get(0).get();
11     final int steps1 = futures.get(1).get();
12     assertThat(Math.abs(steps1 - steps0), lessThanOrEqualTo(value));
13   }
14 
15   public final AtomicBoolean running = new AtomicBoolean(true);
16 
17   /**
18    * {@link DelayHaltCaller}
19    */
20   private final class DelayHaltCaller implements Callable {
21     @Override
22     public Integer call() throws Exception {
23       Thread.sleep(100L);
24       running.compareAndSet(true, false);
25       return 0;
26     }
27   }

设计实现

Track的状态机

消费者, Coordinator, Track 三者之间的协作关系

实现源码

01 import java.util.concurrent.atomic.AtomicBoolean;
02 import java.util.concurrent.atomic.AtomicInteger;
03 
04 /**
05  * {@link Coordinators}
06  *
07  * @author jushi
08  * @created 2010-12-14
09  */
10 public final class Coordinators {
11   private Coordinators() {}
12 
13   /**
14    * Create a new new {@link Coordinator} with valume.
15    *
16    * @param range for coordination.
17    *
18    * @return a new {@link Coordinator} with valume.
19    */
20   public static Coordinator coordinator(final int range) {
21     return new InnerCoordinator(range);
22   }
23 
24   /** {@link InnerCoordinator} */
25   private final static class InnerCoordinator implements Coordinator {
26 
27     public InnerCoordinator(final int range) { this.range = range; }
28 
29     @Override
30     public Track track(final String name) { return new InnerTrack(name); }
31 
32     private void arrived(final InnerTrack track) {
33       arriveds.incrementAndGet();
34       startNextRoundIfAllTrackArrived();
35     }
36 
37     private void decrement() {
38       tracks.decrementAndGet();
39       startNextRoundIfAllTrackArrived();
40     }
41 
42     private boolean flag() { return flag.get(); }
43 
44     private void increment() { tracks.incrementAndGet(); }
45 
46     private void reverseFlag() { flag.set(!flag.get()); }
47 
48     private void startNextRoundIfAllTrackArrived() {
49       final int currentArriveds = arriveds.get();
50       if (currentArriveds < tracks.get()) return;
51       if (!arriveds.compareAndSet(currentArriveds, 0)) return;
52       reverseFlag();
53     }
54 
55     private final int range;
56     private final AtomicInteger arriveds = new AtomicInteger();
57     private final AtomicInteger tracks = new AtomicInteger();
58     private final AtomicBoolean flag = new AtomicBoolean();
59 
60     /** {@link InnerTrack} */
61     private final class InnerTrack implements Track {
62 
63       public InnerTrack(final String name) {
64         this.name = name;
65         init();
66         increment();
67       }
68 
69       @Override
70       public synchronized void destory() {
71         if (destoryed) return;
72         destoryed = true;
73         decrement();
74       }
75 
76       @Override
77       public synchronized boolean move() {
78         if (destoryed)
79           throw new IllegalStateException("Can\'t move a destoryed track");
80         return state.move(this);
81       }
82 
83       private boolean pause() {
84         arrived(this);
85         state = State.WAITTING;
86         return false;
87       }
88 
89       private boolean canResume() { return flag != flag(); }
90 
91       private boolean forward() {
92         steps++;
93         return true;
94       }
95 
96       private boolean await() { return false; }
97 
98       private void init() {
99         steps = 0;
100         flag = flag();
101         state = State.RUNNING;
102       }
103 
104       private boolean isNotArrived() { return steps < range; }
105 
106       private boolean resume() {
107         init();
108         return false;
109       }
110 
111       private final String name;
112 
113       private int steps;
114       private State state;
115       private boolean destoryed;
116       private boolean flag;
117 
118     }
119 
120     /** {@link State} */
121     private enum State {
122       RUNNING {
123         @Override
124         boolean move(final InnerTrack track) {
125           return track.isNotArrived() ? track.forward() : track.pause();
126         }
127       },
128       WAITTING {
129         @Override
130         boolean move(final InnerTrack track) {
131           return track.canResume() ? track.resume() : track.await();
132         }
133       };
134 
135       abstract boolean move(InnerTrack track);
136     }
137 
138   }
139 }

目前TimeTunnel已经在淘蝌蚪上开源, 更多源码请移步https://github.com/huolong/timetunnel.

建议继续学习：

QQ技术交流群：445447336，欢迎加入！
扫一扫订阅我的微信号：IT技术博客大学习

<< 前一篇：ZeroMQ 的模式

后一篇：Amazon分布式系统Dynamo >>

文章信息

作者：聚石来源：淘宝数据平台与产品部官方博客 tbdata.org
标签： TimeTunnel 同步均衡消息分发
发布时间：2011-03-02 22:53:12

建议继续学习

近3天十大热文