问题背景及要求
- 需要对评论进行点赞次数和被评论次数进行统计,或者更多维度
- 要求高并发、高性能计数,允许极端情况丢失一些统计次数,例如宕机
- 评论很多,不能为每一个评论都一直保留其计数器,计数器需要有回收机制
问题抽象及分析
根据以上需求,为了方便编码与测试,我们把需求转化为以下接口
/** * 计数器 */ public interface Counter { /** * 取出统计数据,用Saver去持久化(仅定时器会调用,无并发) * @param saver */ void save(Saver saver); /** * 计数(有并发) * @param key 业务ID * @param like 点赞 * @param comment 评论 */ void add(String key, int like, int comment); /** * 持久化器,将数量持久化到数据库等 */ @FunctionalInterface interface Saver{ void save(String key, int like, int comment); } }
简单分析可知,计数器比较简单,用AtomicInteger便能保证原子性,但考虑到计数器会被回收,则可能会出现这样的场景:某计数器已被回收了,此时继续在该计数器上计数,便会造成数据丢失,因此要处理该并发问题
解决方案
方案一
使用原生锁来解决竞争问题
/** * 直接对所有操作上锁,来保证线程安全 */ public class SynchronizedCounter implements Counter{ private HashMap<String, Adder> map = new HashMap<>(); @Override public synchronized void save(Saver saver) { map.forEach((key, value)->{//因为已加锁,所以可以安全地取数据 saver.save(key, value.like, value.comment); }); map = new HashMap<>(); } @Override public synchronized void add(String key, int like, int comment) { //因为已加锁,所以可以安全地更新数据 Adder adder = map.computeIfAbsent(key, x -> new Adder()); adder.like += like; adder.comment += comment; } static class Adder{ private int like; private int comment; } }
方案点评:该方案让业务线程和定时保存线程竞争同一把实例锁,让他们互斥地访问,解决了竞争问题,但锁粒度太粗爆,性能低下
方案二
为了循序渐进,我们把“计数器需要有回收机制”这条要求去掉,这样我们可以很容易地利用上AtomicInteger这个类
/** * 不回收计数器,问题变得简单许多 */ public class IncompleteCounter implements Counter { private ConcurrentHashMap<String, Adder> map = new ConcurrentHashMap<>(); @Override public void save(Saver saver) { map.forEach((key, value)->{//利用了AtomicInteger的原子特性,可以线程安全地取出所有计数,并置0(因为还会继续使用) saver.save(key, value.like.getAndSet(0), value.comment.getAndSet(0)); }); //因为不回收,所以不用考虑Adder被回收丢弃后,仍被其它线程使用的情况(因为没有锁,所以这种情况是可能发生的) } @Override public void add(String key, int like, int comment) { Adder adder = map.computeIfAbsent(key, k -> new Adder()); adder.like.addAndGet(like);//利用AtomicInteger的原子特性,保证了线程安全 adder.comment.addAndGet(comment); } static class Adder{ AtomicInteger like = new AtomicInteger(); AtomicInteger comment = new AtomicInteger(); } }
方案点评:除了没解决回收问题,简单高效
方案三
因为调用save的线程没有并发情况,阻塞也没关系,经分析可巧妙地使用读写锁,同时又不让add方法进入阻塞
/** * 巧妙地利用读写锁,及save方法可阻塞的特点,实现add操作无阻塞 */ public class ReadWriteLockCounter implements Counter { private volatile MapWithLock mapWithLock = new MapWithLock(); @Override public void save(Saver saver) { MapWithLock preMapWithLock = mapWithLock; mapWithLock = new MapWithLock(); //不会一直阻塞,因为mapWithLock已被替换,新的add调用会拿到新的mapWithLock preMapWithLock.lock.writeLock().lock(); preMapWithLock.map.forEach((key,value)->{ //value已经废弃,故无需value.like.getAndSet(0) saver.save(key, value.like.get(), value.comment.get()); }); //不能释放该锁,否则add方法中,对被替换掉的MapWithLock.lock执行tryLock会成功 //也许,这是你第一次见到的不需要且不允许释放的锁:) } @Override public void add(String key, int like, int comment) { MapWithLock mapWithLock; //如果通过tryLock获取锁失败,则表示该mapWithLock已经被废弃了(因为只有废弃了的MapWithLock才会加写锁),故重新获取最新的mapWithLock while(!(mapWithLock = this.mapWithLock).lock.readLock().tryLock()); try{ Adder adder = mapWithLock.map.computeIfAbsent(key, k -> new Adder()); adder.like.getAndAdd(like); adder.comment.getAndAdd(comment); }finally { mapWithLock.lock.readLock().unlock(); } } static class Adder{ private AtomicInteger like = new AtomicInteger(); private AtomicInteger comment = new AtomicInteger(); } static class MapWithLock{ private ConcurrentHashMap<String, Adder> map = new ConcurrentHashMap<>(); private ReadWriteLock lock = new ReentrantReadWriteLock(); } }
方案点评:减少了锁的粒度,同时add线程可以相互兼容,大幅提升了并发能力,save线程虽会阻塞,但结合其定时执行的特点,并不受影响,且即使极端情况也不会一直阻塞
方案四
使用一个原子的state来替换LockCounter中的ReadWriteLock(因为只使用到了它的部分特性),实现wait-free,获得更高性能
/** * ReadWriteLockCounter的改进版,去掉ReadWriteLock,结合当前场景,实现一个wait-free的简易读写锁<br/> */ public class CustomLockCounter implements Counter { private volatile MapWithState mapWithState = new MapWithState(); @Override public void save(Saver saver) { MapWithState preMapWithState = mapWithState; mapWithState = new MapWithState(); //compareAndSet失败则表示该MapWithState正在被使用,等其使用完,它不会一直失败,因为mapWithState已经被替换 while(!preMapWithState.state.compareAndSet(0,Integer.MIN_VALUE)){ Thread.yield(); } preMapWithState.map.forEach((key, value)->{ //value已经废弃,故无需value.like.getAndSet(0) saver.save(key, value.like.get(), value.comment.get()); }); } @Override public void add(String key, int like, int comment) { MapWithState mapWithState;//add的并发,不可能将Integer.MIN_VALUE自增成正数(设置为Integer.MIN_VALUE时,该MapWithState已经被废弃了) while((mapWithState = this.mapWithState).state.getAndIncrement()<0); try{ Adder adder = mapWithState.map.computeIfAbsent(key, k -> new Adder()); adder.like.getAndAdd(like); adder.comment.getAndAdd(comment); }finally { mapWithState.state.getAndDecrement(); } } static class Adder{ private AtomicInteger like = new AtomicInteger(); private AtomicInteger comment = new AtomicInteger(); } static class MapWithState { private ConcurrentHashMap<String, Adder> map = new ConcurrentHashMap<>(); private AtomicInteger state = new AtomicInteger(); } }
方案点评:保留了前一方案ReadWriteLockCounter的优点,同时结合场景的特点做了些优化,本质就是将CAS失败重试循环替换成了一条fetch-and-add指令,如果不是因为save是低频执行,本方案可能是最高效的了(暂且忽略ConcurrentHashMap等其它可能的优化空间)
方案五
先假定不会发生竞争,然后检测竞争情况,如果发生竞争,则补偿
/** * 乐观地假定不会发生竞争,如果发生了,则尝试进行补偿 */ public class CompensationCounter implements Counter { private ConcurrentHashMap<String, Adder> map = new ConcurrentHashMap<>(); @Override public void save(Saver saver) { for(Iterator<Map.Entry<String, Adder>> it = map.entrySet().iterator(); it.hasNext();){ Map.Entry<String, Adder> entry = it.next(); it.remove(); entry.getValue().discarded = true; saver.save(entry.getKey(), entry.getValue().like.getAndSet(0), entry.getValue().comment.getAndSet(0));//需将计数器置0,此处存在竞争 } } @Override public void add(String key, int like, int comment) { Adder adder = map.computeIfAbsent(key, k -> new Adder()); adder.like.addAndGet(like); adder.comment.addAndGet(comment); if(adder.discarded){//如果数量加在了废弃的Adder上面,则执行补偿逻辑 int likeTemp = adder.like.getAndSet(0); int commentTemp = adder.comment.getAndSet(0); //即使此后又有线程在计数器上计数了也无妨 if(likeTemp != 0 || commentTemp != 0){ add(key, likeTemp, commentTemp);//补偿 }//也可能已经被其它线程取走了,但并不影响业务正确性 } } static class Adder{ AtomicInteger like = new AtomicInteger(); AtomicInteger comment = new AtomicInteger(); volatile boolean discarded = false;//只有保存线程会将它改为true,故使用volatile便能保证线程安全 } }
方案点评:跟乐观锁的思路类似,在竞争激烈的情况下,一般不会有最优性能,但此处因为save方法是低频执行的且自身无并发,add方法才有高并发,故失败补偿其实很少真正被执行,这也是为什么测试结果中本方案性能最优的原因
性能测试
最终我们来测试一下各方案的性能,因为我们抽象出了一个统一的接口,故测试也较为容易
import java.util.Random; import java.util.concurrent.CountDownLatch; import java.util.concurrent.atomic.AtomicInteger; public class CounterTester { private static final int THREAD_SIZE = 6;//add方法的并发线程数 private static final int ADD_SIZE = 5000000;//测试规模 private static final int KEYS_SIZE = 128*1024; public static void main(String[] args) throws InterruptedException { Counter[] counters = new Counter[]{new SynchronizedCounter(), new IncompleteCounter(), new ReadWriteLockCounter(), new CustomLockCounter(), new CompensationCounter()}; String[] keys = new String[KEYS_SIZE]; Random random = new Random(); for (int i = 0; i < keys.length; i++) { keys[i]=String.valueOf(random.nextInt(KEYS_SIZE*1024)); } for (Counter counter : counters) { AtomicInteger totalLike = new AtomicInteger(); AtomicInteger totalComment = new AtomicInteger(); AtomicInteger savedTotalLike = new AtomicInteger(); AtomicInteger savedTotalComment = new AtomicInteger(); Counter.Saver saver = (key, like, comment) -> { savedTotalLike.addAndGet(like);//模拟被持久化到数据库,记录数量以便后续校验正确性 savedTotalComment.addAndGet(comment);//同上 }; CountDownLatch latch = new CountDownLatch(THREAD_SIZE); long start = System.currentTimeMillis(); for (int i = 0; i < THREAD_SIZE; i++) { new Thread(()->{ Random r = new Random(); int like, comment; for (int j = 0; j < ADD_SIZE; j++) { like = 2; comment = 4; counter.add(keys[r.nextInt(KEYS_SIZE)], like, comment); totalLike.addAndGet(like); totalComment.addAndGet(comment); } latch.countDown(); }).start(); } Thread saveThread = new Thread(()->{ while(latch.getCount() != 0){ try { Thread.sleep(100);//模拟100毫秒执行一次持久化 } catch (InterruptedException e) {} counter.save(saver); } counter.save(saver); }); saveThread.start(); latch.await(); System.out.println(counter.getClass().getSimpleName() +" cost:\t"+(System.currentTimeMillis() - start)); saveThread.join(); boolean error = savedTotalLike.get() != totalLike.get() || savedTotalComment.get() != totalComment.get(); (error?System.err:System.out).println("saved:\tlike="+savedTotalLike.get()+"\tcomment="+savedTotalComment.get()); (error?System.err:System.out).println("added:\tlike="+totalLike.get()+"\tcomment="+totalComment.get()+"\n"); } } }
在jdk11(jdk8也基本一致)下的测试结果如下:
注:方案二的IncompleteCounter并未完成回收,仅作对比
SynchronizedCounter cost: 12377 saved: like=60000000 comment=120000000 added: like=60000000 comment=120000000 IncompleteCounter cost: 2560 saved: like=60000000 comment=120000000 added: like=60000000 comment=120000000 ReadWriteLockCounter cost: 7902 saved: like=60000000 comment=120000000 added: like=60000000 comment=120000000 CustomLockCounter cost: 3541 saved: like=60000000 comment=120000000 added: like=60000000 comment=120000000 CompensationCounter cost: 2093 saved: like=60000000 comment=120000000 added: like=60000000 comment=120000000
小结
非阻塞同步算法一般不需要我们去设计,直接使用现有的工具便可,但如果真想通过它进一步去压榨性能,应细心分析各线程穿插执行的情况,同时结合业务场景来考虑(也许在A场景不允许的情况,在B场景是允许的)
原文地址:https://www.cnblogs.com/trytocatch/p/counter-persistence.html