概要
Caffeine是一个高性能,高命中率,低内存占用,near optimal 的本地缓存,简单来说它是Guava Cache的优化加强版,有些文章把Caffeine称为“新一代的缓存”、“现代缓存之王”。本文将重点讲解Caffeine的高性能设计,以及对应部分的源码分析。
与Guava Cache比较
如果你对Guava Cache还不理解的话,可以点击这里来看一下我之前写过关于Guava Cache的文章。
大家都知道,Spring5即将放弃掉Guava Cache作为缓存机制,而改用Caffeine作为新的本地Cache的组件,这对于Caffeine来说是一个很大的肯定。为什么Spring会这样做呢?其实在Caffeine的Benchmarks里给出了好靓仔的数据,对读和写的场景,还有跟其他几个缓存工具进行了比较,Caffeine的性能都表现很突出。
使用Caffeine
Caffeine为了方便大家使用以及从Guava Cache切换过来(很有针对性啊~),借鉴了Guava Cache大部分的概念(诸如核心概念Cache、LoadingCache、CacheLoader、CacheBuilder等等),对于Caffeine的理解只要把它当作Guava Cache就可以了。
使用上,大家只要把Caffeine的包引进来,然后换一下cache的实现类,基本应该就没问题了。这对与已经使用过Guava Cache的同学来说没有任何难度,甚至还有一点熟悉的味道,如果你之前没有使用过Guava Cache,可以查看Caffeine的官方API说明文档,其中Population,Eviction,Removal,Refresh,Statistics,Cleanup,Policy等等这些特性都是跟Guava Cache基本一样的。
下面给出一个例子说明怎样创建一个Cache:
private static LoadingCache<String, String> cache = Caffeine.newBuilder() //最大个数限制 .maximumSize(256L) //初始化容量 .initialCapacity(1) //访问后过期(包括读和写) .expireAfterAccess(2, TimeUnit.DAYS) //写后过期 .expireAfterWrite(2, TimeUnit.HOURS) //写后自动异步刷新 .refreshAfterWrite(1, TimeUnit.HOURS) //记录下缓存的一些统计数据,例如命中率等 .recordStats() //cache对缓存写的通知回调 .writer(new CacheWriter<Object, Object>() { @Override public void write(@NonNull Object key, @NonNull Object value) { log.info("key={}, CacheWriter write", key); } @Override public void delete(@NonNull Object key, @Nullable Object value, @NonNull RemovalCause cause) { log.info("key={}, cause={}, CacheWriter delete", key, cause); } }) //使用CacheLoader创建一个LoadingCache .build(new CacheLoader<String, String>() { //同步加载数据 @Nullable @Override public String load(@NonNull String key) throws Exception { return "value_" + key; } //异步加载数据 @Nullable @Override public String reload(@NonNull String key, @NonNull String oldValue) throws Exception { return "value_" + key; } });
Caffeine的高性能设计
判断一个缓存的好坏最核心的指标就是命中率,影响缓存命中率有很多因素,包括业务场景、淘汰策略、清理策略、缓存容量等等。如果作为本地缓存, 它的性能的情况,资源的占用也都是一个很重要的指标。下面
我们来看看Caffeine在这几个方面是怎么着手的,如何做优化的。
(注:本文不会分析Caffeine全部源码,只会对核心设计的实现进行分析,但我建议读者把Caffeine的源码都涉猎一下,有个overview才能更好理解本文。如果你看过Guava Cache的源码也行,代码的数据结构和处理逻辑很类似的。源码基于:caffeine-2.8.0.jar)
W-TinyLFU整体设计
上面说到淘汰策略是影响缓存命中率的因素之一,一般比较简单的缓存就会直接用到LFU(Least Frequently Used,即最不经常使用)或者LRU(Least Recently Used,即最近最少使用),而Caffeine就是使用了W-TinyLFU算法。
W-TinyLFU看名字就能大概猜出来,它是LFU的变种,也是一种缓存淘汰算法。那为什么要使用W-TinyLFU呢?
LRU和LFU的缺点
- LRU实现简单,在一般情况下能够表现出很好的命中率,是一个“性价比”很高的算法,平时也很常用。虽然LRU对突发性的稀疏流量(sparse bursts)表现很好,但同时也会产生缓存污染,举例来说,如果偶然性的要对全量数据进行遍历,那么“历史访问记录”就会被刷走,造成污染。
- 如果数据的分布在一段时间内是固定的话,那么LFU可以达到最高的命中率。但是LFU有两个缺点,第一,它需要给每个记录项维护频率信息,每次访问都需要更新,这是个巨大的开销;第二,对突发性的稀疏流量无力,因为前期经常访问的记录已经占用了缓存,偶然的流量不太可能会被保留下来,而且过去的一些大量被访问的记录在将来也不一定会使用上,这样就一直把“坑”占着了。
无论LRU还是LFU都有其各自的缺点,不过,现在已经有很多针对其缺点而改良、优化出来的变种算法。
TinyLFU
TinyLFU就是其中一个优化算法,它是专门为了解决LFU上述提到的两个问题而被设计出来的。
解决第一个问题是采用了Count–Min Sketch算法。
解决第二个问题是让记录尽量保持相对的“新鲜”(Freshness Mechanism),并且当有新的记录插入时,可以让它跟老的记录进行“PK”,输者就会被淘汰,这样一些老的、不再需要的记录就会被剔除。
下图是TinyLFU设计图(来自官方)
统计频率Count–Min Sketch算法
如何对一个key进行统计,但又可以节省空间呢?(不是简单的使用HashMap,这太消耗内存了),注意哦,不需要精确的统计,只需要一个近似值就可以了,怎么样,这样场景是不是很熟悉,如果你是老司机,或许已经联想到布隆过滤器(Bloom Filter)的应用了。
没错,将要介绍的Count–Min Sketch的原理跟Bloom Filter一样,只不过Bloom Filter只有0和1的值,那么你可以把Count–Min Sketch看作是“数值”版的Bloom Filter。
更多关于Count–Min Sketch的介绍请自行搜索。
在TinyLFU中,近似频率的统计如下图所示:
对一个key进行多次hash函数后,index到多个数组位置后进行累加,查询时取多个值中的最小值即可。
Caffeine对这个算法的实现在FrequencySketch
类。但Caffeine对此有进一步的优化,例如Count–Min
Sketch使用了二维数组,Caffeine只是用了一个一维的数组;再者,如果是数值类型的话,这个数需要用int或long来存储,但是Caffeine认为缓存的访问频率不需要用到那么大,只需要15就足够,一般认为达到15次的频率算是很高的了,而且Caffeine还有另外一个机制来使得这个频率进行衰退减半(下面就会讲到)。如果最大是15的话,那么只需要4个bit就可以满足了,一个long有64bit,可以存储16个这样的统计数,Caffeine就是这样的设计,使得存储效率提高了16倍。
Caffeine对缓存的读写(afterRead
和afterWrite
方法)都会调用onAccess
s方法,而onAccess方法里有一句:
frequencySketch().increment(key);
这句就是追加记录的频率,下面我们看看具体实现
//FrequencySketch的一些属性 //种子数 static final long[] SEED = { // A mixture of seeds from FNV-1a, CityHash, and Murmur3 0xc3a5c85c97cb3127L, 0xb492b66fbe98f273L, 0x9ae16a3b2f90404fL, 0xcbf29ce484222325L}; static final long RESET_MASK = 0x7777777777777777L; static final long ONE_MASK = 0x1111111111111111L; int sampleSize; //为了快速根据hash值得到table的index值的掩码 //table的长度size一般为2的n次方,而tableMask为size-1,这样就可以通过&操作来模拟取余操作,速度快很多,老司机都知道 int tableMask; //存储数据的一维long数组 long[] table; int size; /** * Increments the popularity of the element if it does not exceed the maximum (15). The popularity * of all elements will be periodically down sampled when the observed events exceeds a threshold. * This process provides a frequency aging to allow expired long term entries to fade away. * * @param e the element to add */ public void increment(@NonNull E e) { if (isNotInitialized()) { return; } //根据key的hashCode通过一个哈希函数得到一个hash值 //本来就是hashCode了,为什么还要再做一次hash?怕原来的hashCode不够均匀分散,再打散一下。 int hash = spread(e.hashCode()); //这句光看有点难理解 //就如我刚才说的,Caffeine把一个long的64bit划分成16个等分,每一等分4个bit。 //这个start就是用来定位到是哪一个等分的,用hash值低两位作为随机数,再左移2位,得到一个小于16的值 int start = (hash & 3) << 2; //indexOf方法的意思就是,根据hash值和不同种子得到table的下标index //这里通过四个不同的种子,得到四个不同的下标index int index0 = indexOf(hash, 0); int index1 = indexOf(hash, 1); int index2 = indexOf(hash, 2); int index3 = indexOf(hash, 3); //根据index和start(+1, +2, +3)的值,把table[index]对应的等分追加1 //这个incrementAt方法有点难理解,看我下面的解释 boolean added = incrementAt(index0, start); added |= incrementAt(index1, start + 1); added |= incrementAt(index2, start + 2); added |= incrementAt(index3, start + 3); //这个reset等下说 if (added && (++size == sampleSize)) { reset(); } } /** * Increments the specified counter by 1 if it is not already at the maximum value (15). * * @param i the table index (16 counters) * @param j the counter to increment * @return if incremented */ boolean incrementAt(int i, int j) { //这个j表示16个等分的下标,那么offset就是相当于在64位中的下标(这个自己想想) int offset = j << 2; //上面提到Caffeine把频率统计最大定为15,即0xfL //mask就是在64位中的掩码,即1111后面跟很多个0 long mask = (0xfL << offset); //如果&的结果不等于15,那么就追加1。等于15就不会再加了 if ((table[i] & mask) != mask) { table[i] += (1L << offset); return true; } return false; } /** * Returns the table index for the counter at the specified depth. * * @param item the element's hash * @param i the counter depth * @return the table index */ int indexOf(int item, int i) { long hash = SEED[i] * item; hash += hash >>> 32; return ((int) hash) & tableMask; } /** * Applies a supplemental hash function to a given hashCode, which defends against poor quality * hash functions. */ int spread(int x) { x = ((x >>> 16) ^ x) * 0x45d9f3b; x = ((x >>> 16) ^ x) * 0x45d9f3b; return (x >>> 16) ^ x; }
知道了追加方法,那么读取方法frequency
就很容易理解了。
/** * Returns the estimated number of occurrences of an element, up to the maximum (15). * * @param e the element to count occurrences of * @return the estimated number of occurrences of the element; possibly zero but never negative */ @NonNegative public int frequency(@NonNull E e) { if (isNotInitialized()) { return 0; } //得到hash值,跟上面一样 int hash = spread(e.hashCode()); //得到等分的下标,跟上面一样 int start = (hash & 3) << 2; int frequency = Integer.MAX_VALUE; //循环四次,分别获取在table数组中不同的下标位置 for (int i = 0; i < 4; i++) { int index = indexOf(hash, i); //这个操作就不多说了,其实跟上面incrementAt是一样的,定位到table[index] + 等分的位置,再根据mask取出计数值 int count = (int) ((table[index] >>> ((start + i) << 2)) & 0xfL); //取四个中的较小值 frequency = Math.min(frequency, count); } return frequency; }
通过代码和注释或者读者可能难以理解,下图是我画出来帮助大家理解的结构图。
注意紫色虚线框,其中蓝色小格就是需要计算的位置:
保新机制
为了让缓存保持“新鲜”,剔除掉过往频率很高但之后不经常的缓存,Caffeine有一个Freshness Mechanism。做法很简答,就是当整体的统计计数(当前所有记录的频率统计之和,这个数值内部维护)达到某一个值时,那么所有记录的频率统计除以2。
从上面的代码
//size变量就是所有记录的频率统计之,即每个记录加1,这个size都会加1 //sampleSize一个阈值,从FrequencySketch初始化可以看到它的值为maximumSize的10倍 if (added && (++size == sampleSize)) { reset(); }
看到reset
方法就是做这个事情
/** Reduces every counter by half of its original value. */ void reset() { int count = 0; for (int i = 0; i < table.length; i++) { count += Long.bitCount(table[i] & ONE_MASK); table[i] = (table[i] >>> 1) & RESET_MASK; } size = (size >>> 1) - (count >>> 2); }
关于这个reset方法,为什么是除以2,而不是其他,及其正确性,在最下面的参考资料的TinyLFU论文中3.3章节给出了数学证明,大家有兴趣可以看看。
增加一个Window?
Caffeine通过测试发现TinyLFU在面对突发性的稀疏流量(sparse bursts)时表现很差,因为新的记录(new items)还没来得及建立足够的频率就被剔除出去了,这就使得命中率下降。
于是Caffeine设计出一种新的policy,即Window Tiny LFU(W-TinyLFU),并通过实验和实践发现W-TinyLFU比TinyLFU表现的更好。
W-TinyLFU的设计如下所示(两图等价):
它主要包括两个缓存模块,主缓存是SLRU(Segmented LRU,即分段LRU),SLRU包括一个名为protected和一个名为probation的缓存区。通过增加一个缓存区(即Window Cache),当有新的记录插入时,会先在window区呆一下,就可以避免上述说的sparse bursts问题。
淘汰策略(eviction policy)
当window区满了,就会根据LRU把candidate(即淘汰出来的元素)放到probation区,如果probation区也满了,就把candidate和probation将要淘汰的元素victim,两个进行“PK”,胜者留在probation,输者就要被淘汰了。
而且经过实验发现当window区配置为总容量的1%,剩余的99%当中的80%分给protected区,20%分给probation区时,这时整体性能和命中率表现得最好,所以Caffeine默认的比例设置就是这个。
不过这个比例Caffeine会在运行时根据统计数据(statistics)去动态调整,如果你的应用程序的缓存随着时间变化比较快的话,那么增加window区的比例可以提高命中率,相反缓存都是比较固定不变的话,增加Main Cache区(protected区 +probation区)的比例会有较好的效果。
下面我们看看上面说到的淘汰策略是怎么实现的:
一般缓存对读写操作后都有后续的一系列“维护”操作,Caffeine也不例外,这些操作都在maintenance
方法,我们将要说到的淘汰策略也在里面。
这方法比较重要,下面也会提到,所以这里只先说跟“淘汰策略”有关的evictEntries
和climb
。
/** * Performs the pending maintenance work and sets the state flags during processing to avoid * excess scheduling attempts. The read buffer, write buffer, and reference queues are * drained, followed by expiration, and size-based eviction. * * @param task an additional pending task to run, or {@code null} if not present */ @GuardedBy("evictionLock") void maintenance(@Nullable Runnable task) { lazySetDrainStatus(PROCESSING_TO_IDLE); try { drainReadBuffer(); drainWriteBuffer(); if (task != null) { task.run(); } drainKeyReferences(); drainValueReferences(); expireEntries(); //把符合条件的记录淘汰掉 evictEntries(); //动态调整window区和protected区的大小 climb(); } finally { if ((drainStatus() != PROCESSING_TO_IDLE) || !casDrainStatus(PROCESSING_TO_IDLE, IDLE)) { lazySetDrainStatus(REQUIRED); } } }
先说一下Caffeine对上面说到的W-TinyLFU策略的实现用到的数据结构:
//最大的个数限制 long maximum; //当前的个数 long weightedSize; //window区的最大限制 long windowMaximum; //window区当前的个数 long windowWeightedSize; //protected区的最大限制 long mainProtectedMaximum; //protected区当前的个数 long mainProtectedWeightedSize; //下一次需要调整的大小(还需要进一步计算) double stepSize; //window区需要调整的大小 long adjustment; //命中计数 int hitsInSample; //不命中的计数 int missesInSample; //上一次的缓存命中率 double previousSampleHitRate; final FrequencySketch<K> sketch; //window区的LRU queue(FIFO) final AccessOrderDeque<Node<K, V>> accessOrderWindowDeque; //probation区的LRU queue(FIFO) final AccessOrderDeque<Node<K, V>> accessOrderProbationDeque; //protected区的LRU queue(FIFO) final AccessOrderDeque<Node<K, V>> accessOrderProtectedDeque;
以及默认比例设置(意思看注释)
/** The initial percent of the maximum weighted capacity dedicated to the main space. */ static final double PERCENT_MAIN = 0.99d; /** The percent of the maximum weighted capacity dedicated to the main's protected space. */ static final double PERCENT_MAIN_PROTECTED = 0.80d; /** The difference in hit rates that restarts the climber. */ static final double HILL_CLIMBER_RESTART_THRESHOLD = 0.05d; /** The percent of the total size to adapt the window by. */ static final double HILL_CLIMBER_STEP_PERCENT = 0.0625d; /** The rate to decrease the step size to adapt by. */ static final double HILL_CLIMBER_STEP_DECAY_RATE = 0.98d; /** The maximum number of entries that can be transfered between queues. */
重点来了,evictEntries和climb方法:
/** Evicts entries if the cache exceeds the maximum. */ @GuardedBy("evictionLock") void evictEntries() { if (!evicts()) { return; } //淘汰window区的记录 int candidates = evictFromWindow(); //淘汰Main区的记录 evictFromMain(candidates); } /** * Evicts entries from the window space into the main space while the window size exceeds a * maximum. * * @return the number of candidate entries evicted from the window space */ //根据W-TinyLFU,新的数据都会无条件的加到admission window //但是window是有大小限制,所以要“定期”做一下“维护” @GuardedBy("evictionLock") int evictFromWindow() { int candidates = 0; //查看window queue的头部节点 Node<K, V> node = accessOrderWindowDeque().peek(); //如果window区超过了最大的限制,那么就要把“多出来”的记录做处理 while (windowWeightedSize() > windowMaximum()) { // The pending operations will adjust the size to reflect the correct weight if (node == null) { break; } //下一个节点 Node<K, V> next = node.getNextInAccessOrder(); if (node.getWeight() != 0) { //把node定位在probation区 node.makeMainProbation(); //从window区去掉 accessOrderWindowDeque().remove(node); //加入到probation queue,相当于把节点移动到probation区(晋升了) accessOrderProbationDeque().add(node); candidates++; //因为移除了一个节点,所以需要调整window的size setWindowWeightedSize(windowWeightedSize() - node.getPolicyWeight()); } //处理下一个节点 node = next; } return candidates; }
evictFromMain
方法:
/** * Evicts entries from the main space if the cache exceeds the maximum capacity. The main space * determines whether admitting an entry (coming from the window space) is preferable to retaining * the eviction policy's victim. This is decision is made using a frequency filter so that the * least frequently used entry is removed. * * The window space candidates were previously placed in the MRU position and the eviction * policy's victim is at the LRU position. The two ends of the queue are evaluated while an * eviction is required. The number of remaining candidates is provided and decremented on * eviction, so that when there are no more candidates the victim is evicted. * * @param candidates the number of candidate entries evicted from the window space */ //根据W-TinyLFU,从window晋升过来的要跟probation区的进行“PK”,胜者才能留下 @GuardedBy("evictionLock") void evictFromMain(int candidates) { int victimQueue = PROBATION; //victim是probation queue的头部 Node<K, V> victim = accessOrderProbationDeque().peekFirst(); //candidate是probation queue的尾部,也就是刚从window晋升来的 Node<K, V> candidate = accessOrderProbationDeque().peekLast(); //当cache不够容量时才做处理 while (weightedSize() > maximum()) { // Stop trying to evict candidates and always prefer the victim if (candidates == 0) { candidate = null; } //对candidate为null且victim为bull的处理 if ((candidate == null) && (victim == null)) { if (victimQueue == PROBATION) { victim = accessOrderProtectedDeque().peekFirst(); victimQueue = PROTECTED; continue; } else if (victimQueue == PROTECTED) { victim = accessOrderWindowDeque().peekFirst(); victimQueue = WINDOW; continue; } // The pending operations will adjust the size to reflect the correct weight break; } //对节点的weight为0的处理 if ((victim != null) && (victim.getPolicyWeight() == 0)) { victim = victim.getNextInAccessOrder(); continue; } else if ((candidate != null) && (candidate.getPolicyWeight() == 0)) { candidate = candidate.getPreviousInAccessOrder(); candidates--; continue; } // Evict immediately if only one of the entries is present if (victim == null) { @SuppressWarnings("NullAway") Node<K, V> previous = candidate.getPreviousInAccessOrder(); Node<K, V> evict = candidate; candidate = previous; candidates--; evictEntry(evict, RemovalCause.SIZE, 0L); continue; } else if (candidate == null) { Node<K, V> evict = victim; victim = victim.getNextInAccessOrder(); evictEntry(evict, RemovalCause.SIZE, 0L); continue; } // Evict immediately if an entry was collected K victimKey = victim.getKey(); K candidateKey = candidate.getKey(); if (victimKey == null) { @NonNull Node<K, V> evict = victim; victim = victim.getNextInAccessOrder(); evictEntry(evict, RemovalCause.COLLECTED, 0L); continue; } else if (candidateKey == null) { candidates--; @NonNull Node<K, V> evict = candidate; candidate = candidate.getPreviousInAccessOrder(); evictEntry(evict, RemovalCause.COLLECTED, 0L); continue; } //放不下的节点直接处理掉 if (candidate.getPolicyWeight() > maximum()) { candidates--; Node<K, V> evict = candidate; candidate = candidate.getPreviousInAccessOrder(); evictEntry(evict, RemovalCause.SIZE, 0L); continue; } //根据节点的统计频率frequency来做比较,看看要处理掉victim还是candidate //admit是具体的比较规则,看下面 candidates--; //如果candidate胜出则淘汰victim if (admit(candidateKey, victimKey)) { Node<K, V> evict = victim; victim = victim.getNextInAccessOrder(); evictEntry(evict, RemovalCause.SIZE, 0L); candidate = candidate.getPreviousInAccessOrder(); } else { //如果是victim胜出,则淘汰candidate Node<K, V> evict = candidate; candidate = candidate.getPreviousInAccessOrder(); evictEntry(evict, RemovalCause.SIZE, 0L); } } } /** * Determines if the candidate should be accepted into the main space, as determined by its * frequency relative to the victim. A small amount of randomness is used to protect against hash * collision attacks, where the victim's frequency is artificially raised so that no new entries * are admitted. * * @param candidateKey the key for the entry being proposed for long term retention * @param victimKey the key for the entry chosen by the eviction policy for replacement * @return if the candidate should be admitted and the victim ejected */ @GuardedBy("evictionLock") boolean admit(K candidateKey, K victimKey) { //分别获取victim和candidate的统计频率 //frequency这个方法的原理和实现上面已经解释了 int victimFreq = frequencySketch().frequency(victimKey); int candidateFreq = frequencySketch().frequency(candidateKey); //谁大谁赢 if (candidateFreq > victimFreq) { return true; //如果相等,candidate小于5都当输了 } else if (candidateFreq <= 5) { // The maximum frequency is 15 and halved to 7 after a reset to age the history. An attack // exploits that a hot candidate is rejected in favor of a hot victim. The threshold of a warm // candidate reduces the number of random acceptances to minimize the impact on the hit rate. return false; } //如果相等且candidate大于5,则随机淘汰一个 int random = ThreadLocalRandom.current().nextInt(); return ((random & 127) == 0); }
climb
方法主要是用来调整window size的,使得Caffeine可以适应你的应用类型(如OLAP或OLTP)表现出最佳的命中率。
下图是官方测试的数据:
我们看看window size的调整是怎么实现的。
调整时用到的默认比例数据:
//与上次命中率之差的阈值 static final double HILL_CLIMBER_RESTART_THRESHOLD = 0.05d; //步长(调整)的大小(跟最大值maximum的比例) static final double HILL_CLIMBER_STEP_PERCENT = 0.0625d; //步长的衰减比例 static final double HILL_CLIMBER_STEP_DECAY_RATE = 0.98d; /** Adapts the eviction policy to towards the optimal recency / frequency configuration. */ //climb方法的主要作用就是动态调整window区的大小(相应的,main区的大小也会发生变化,两个之和为100%)。 //因为区域的大小发生了变化,那么区域内的数据也可能需要发生相应的移动。 @GuardedBy("evictionLock") void climb() { if (!evicts()) { return; } //确定window需要调整的大小 determineAdjustment(); //如果protected区有溢出,把溢出部分移动到probation区。因为下面的操作有可能需要调整到protected区。 demoteFromMainProtected(); long amount = adjustment(); if (amount == 0) { return; } else if (amount > 0) { //增加window的大小 increaseWindow(); } else { //减少window的大小 decreaseWindow(); } }
下面分别展开每个方法来解释:
/** Calculates the amount to adapt the window by and sets {@link #adjustment()} accordingly. */ @GuardedBy("evictionLock") void determineAdjustment() { //如果frequencySketch还没初始化,则返回 if (frequencySketch().isNotInitialized()) { setPreviousSampleHitRate(0.0); setMissesInSample(0); setHitsInSample(0); return; } //总请求量 = 命中 + miss int requestCount = hitsInSample() + missesInSample(); //没达到sampleSize则返回 //默认下sampleSize = 10 * maximum。用sampleSize来判断缓存是否足够”热“。 if (requestCount < frequencySketch().sampleSize) { return; } //命中率的公式 = 命中 / 总请求 double hitRate = (double) hitsInSample() / requestCount; //命中率的差值 double hitRateChange = hitRate - previousSampleHitRate(); //本次调整的大小,是由命中率的差值和上次的stepSize决定的 double amount = (hitRateChange >= 0) ? stepSize() : -stepSize(); //下次的调整大小:如果命中率的之差大于0.05,则重置为0.065 * maximum,否则按照0.98来进行衰减 double nextStepSize = (Math.abs(hitRateChange) >= HILL_CLIMBER_RESTART_THRESHOLD) ? HILL_CLIMBER_STEP_PERCENT * maximum() * (amount >= 0 ? 1 : -1) : HILL_CLIMBER_STEP_DECAY_RATE * amount; setPreviousSampleHitRate(hitRate); setAdjustment((long) amount); setStepSize(nextStepSize); setMissesInSample(0); setHitsInSample(0); } /** Transfers the nodes from the protected to the probation region if it exceeds the maximum. */ //这个方法比较简单,减少protected区溢出的部分 @GuardedBy("evictionLock") void demoteFromMainProtected() { long mainProtectedMaximum = mainProtectedMaximum(); long mainProtectedWeightedSize = mainProtectedWeightedSize(); if (mainProtectedWeightedSize <= mainProtectedMaximum) { return; } for (int i = 0; i < QUEUE_TRANSFER_THRESHOLD; i++) { if (mainProtectedWeightedSize <= mainProtectedMaximum) { break; } Node<K, V> demoted = accessOrderProtectedDeque().poll(); if (demoted == null) { break; } demoted.makeMainProbation(); accessOrderProbationDeque().add(demoted); mainProtectedWeightedSize -= demoted.getPolicyWeight(); } setMainProtectedWeightedSize(mainProtectedWeightedSize); } /** * Increases the size of the admission window by shrinking the portion allocated to the main * space. As the main space is partitioned into probation and protected regions (80% / 20%), for * simplicity only the protected is reduced. If the regions exceed their maximums, this may cause * protected items to be demoted to the probation region and probation items to be demoted to the * admission window. */ //增加window区的大小,这个方法比较简单,思路就像我上面说的 @GuardedBy("evictionLock") void increaseWindow() { if (mainProtectedMaximum() == 0) { return; } long quota = Math.min(adjustment(), mainProtectedMaximum()); setMainProtectedMaximum(mainProtectedMaximum() - quota); setWindowMaximum(windowMaximum() + quota); demoteFromMainProtected(); for (int i = 0; i < QUEUE_TRANSFER_THRESHOLD; i++) { Node<K, V> candidate = accessOrderProbationDeque().peek(); boolean probation = true; if ((candidate == null) || (quota < candidate.getPolicyWeight())) { candidate = accessOrderProtectedDeque().peek(); probation = false; } if (candidate == null) { break; } int weight = candidate.getPolicyWeight(); if (quota < weight) { break; } quota -= weight; if (probation) { accessOrderProbationDeque().remove(candidate); } else { setMainProtectedWeightedSize(mainProtectedWeightedSize() - weight); accessOrderProtectedDeque().remove(candidate); } setWindowWeightedSize(windowWeightedSize() + weight); accessOrderWindowDeque().add(candidate); candidate.makeWindow(); } setMainProtectedMaximum(mainProtectedMaximum() + quota); setWindowMaximum(windowMaximum() - quota); setAdjustment(quota); } /** Decreases the size of the admission window and increases the main's protected region. */ //同上increaseWindow差不多,反操作 @GuardedBy("evictionLock") void decreaseWindow() { if (windowMaximum() <= 1) { return; } long quota = Math.min(-adjustment(), Math.max(0, windowMaximum() - 1)); setMainProtectedMaximum(mainProtectedMaximum() + quota); setWindowMaximum(windowMaximum() - quota); for (int i = 0; i < QUEUE_TRANSFER_THRESHOLD; i++) { Node<K, V> candidate = accessOrderWindowDeque().peek(); if (candidate == null) { break; } int weight = candidate.getPolicyWeight(); if (quota < weight) { break; } quota -= weight; setMainProtectedWeightedSize(mainProtectedWeightedSize() + weight); setWindowWeightedSize(windowWeightedSize() - weight); accessOrderWindowDeque().remove(candidate); accessOrderProbationDeque().add(candidate); candidate.makeMainProbation(); } setMainProtectedMaximum(mainProtectedMaximum() - quota); setWindowMaximum(windowMaximum() + quota); setAdjustment(-quota); }
以上,是Caffeine的W-TinyLFU策略的设计原理及代码实现解析。
异步的高性能读写
一般的缓存每次对数据处理完之后(读的话,已经存在则直接返回,不存在则load数据,保存,再返回;写的话,则直接插入或更新),但是因为要维护一些淘汰策略,则需要一些额外的操作,诸如:
- 计算和比较数据的是否过期
- 统计频率(像LFU或其变种)
- 维护read queue和write queue
- 淘汰符合条件的数据
- 等等。。。
这种数据的读写伴随着缓存状态的变更,Guava Cache的做法是把这些操作和读写操作放在一起,在一个同步加锁的操作中完成,虽然Guava Cache巧妙地利用了JDK的ConcurrentHashMap(分段锁或者无锁CAS)来降低锁的密度,达到提高并发度的目的。但是,对于一些热点数据,这种做法还是避免不了频繁的锁竞争。Caffeine借鉴了数据库系统的WAL(Write-Ahead Logging)思想,即先写日志再执行操作,这种思想同样适合缓存的,执行读写操作时,先把操作记录在缓冲区,然后在合适的时机异步、批量地执行缓冲区中的内容。但在执行缓冲区的内容时,也是需要在缓冲区加上同步锁的,不然存在并发问题,只不过这样就可以把对锁的竞争从缓存数据转移到对缓冲区上。
ReadBuffer
在Caffeine的内部实现中,为了很好的支持不同的Features(如Eviction,Removal,Refresh,Statistics,Cleanup,Policy等等),扩展了很多子类,它们共同的父类是BoundedLocalCache
,而readBuffer
就是作为它们共有的属性,即都是用一样的readBuffer,看定义:
final Buffer<Node<K, V>> readBuffer; readBuffer = evicts() || collectKeys() || collectValues() || expiresAfterAccess() ? new BoundedBuffer<>() : Buffer.disabled();
上面提到Caffeine对每次缓存的读操作都会触发afterRead
/** * Performs the post-processing work required after a read. * * @param node the entry in the page replacement policy * @param now the current time, in nanoseconds * @param recordHit if the hit count should be incremented */ void afterRead(Node<K, V> node, long now, boolean recordHit) { if (recordHit) { statsCounter().recordHits(1); } //把记录加入到readBuffer //判断是否需要立即处理readBuffer //注意这里无论offer是否成功都可以走下去的,即允许写入readBuffer丢失,因为这个 boolean delayable = skipReadBuffer() || (readBuffer.offer(node) != Buffer.FULL); if (shouldDrainBuffers(delayable)) { scheduleDrainBuffers(); } refreshIfNeeded(node, now); } /** * Returns whether maintenance work is needed. * * @param delayable if draining the read buffer can be delayed */ //caffeine用了一组状态来定义和管理“维护”的过程 boolean shouldDrainBuffers(boolean delayable) { switch (drainStatus()) { case IDLE: return !delayable; case REQUIRED: return true; case PROCESSING_TO_IDLE: case PROCESSING_TO_REQUIRED: return false; default: throw new IllegalStateException(); } }
重点看BoundedBuffer
/** * A striped, non-blocking, bounded buffer. * * @author ben.manes@gmail.com (Ben Manes) * @param <E> the type of elements maintained by this buffer */ final class BoundedBuffer<E> extends StripedBuffer<E>
它是一个striped、非阻塞、有界限的buffer,继承于StripedBuffer
类。下面看看StripedBuffer的实现:
/** * A base class providing the mechanics for supporting dynamic striping of bounded buffers. This * implementation is an adaption of the numeric 64-bit {@link java.util.concurrent.atomic.Striped64} * class, which is used by atomic counters. The approach was modified to lazily grow an array of * buffers in order to minimize memory usage for caches that are not heavily contended on. * * @author dl@cs.oswego.edu (Doug Lea) * @author ben.manes@gmail.com (Ben Manes) */ abstract class StripedBuffer<E> implements Buffer<E>
这个StripedBuffer设计的思想是跟Striped64
类似的,通过扩展结构把竞争热点分离。
具体实现是这样的,StripedBuffer维护一个Buffer[]数组,每个元素就是一个RingBuffer
,每个线程用自己threadLocalRandomProbe
属性作为hash值,这样就相当于每个线程都有自己“专属”的RingBuffer,就不会产生竞争啦,而不是用key的hashCode作为hash值,因为会产生热点数据问题。
看看StripedBuffer的属性
/** Table of buffers. When non-null, size is a power of 2. */ //RingBuffer数组 transient volatile Buffer<E> @Nullable[] table; //当进行resize时,需要整个table锁住。tableBusy作为CAS的标记。 static final long TABLE_BUSY = UnsafeAccess.objectFieldOffset(StripedBuffer.class, "tableBusy"); static final long PROBE = UnsafeAccess.objectFieldOffset(Thread.class, "threadLocalRandomProbe"); /** Number of CPUS. */ static final int NCPU = Runtime.getRuntime().availableProcessors(); /** The bound on the table size. */ //table最大size static final int MAXIMUM_TABLE_SIZE = 4 * ceilingNextPowerOfTwo(NCPU); /** The maximum number of attempts when trying to expand the table. */ //如果发生竞争时(CAS失败)的尝试次数 static final int ATTEMPTS = 3; /** Table of buffers. When non-null, size is a power of 2. */ //核心数据结构 transient volatile Buffer<E> @Nullable[] table; /** Spinlock (locked via CAS) used when resizing and/or creating Buffers. */ transient volatile int tableBusy; /** CASes the tableBusy field from 0 to 1 to acquire lock. */ final boolean casTableBusy() { return UnsafeAccess.UNSAFE.compareAndSwapInt(this, TABLE_BUSY, 0, 1); } /** * Returns the probe value for the current thread. Duplicated from ThreadLocalRandom because of * packaging restrictions. */ static final int getProbe() { return UnsafeAccess.UNSAFE.getInt(Thread.currentThread(), PROBE); }
offer
方法,当没初始化或存在竞争时,则扩容为2倍。
实际是调用RingBuffer
的offer方法,把数据追加到RingBuffer后面。
@Override public int offer(E e) { int mask; int result = 0; Buffer<E> buffer; //是否不存在竞争 boolean uncontended = true; Buffer<E>[] buffers = table //是否已经初始化 if ((buffers == null) || (mask = buffers.length - 1) < 0 //用thread的随机值作为hash值,得到对应位置的RingBuffer || (buffer = buffers[getProbe() & mask]) == null //检查追加到RingBuffer是否成功 || !(uncontended = ((result = buffer.offer(e)) != Buffer.FAILED))) { //其中一个符合条件则进行扩容 expandOrRetry(e, uncontended); } return result; } /** * Handles cases of updates involving initialization, resizing, creating new Buffers, and/or * contention. See above for explanation. This method suffers the usual non-modularity problems of * optimistic retry code, relying on rechecked sets of reads. * * @param e the element to add * @param wasUncontended false if CAS failed before call */ //这个方法比较长,但思路还是相对清晰的。 @SuppressWarnings("PMD.ConfusingTernary") final void expandOrRetry(E e, boolean wasUncontended) { int h; if ((h = getProbe()) == 0) { ThreadLocalRandom.current(); // force initialization h = getProbe(); wasUncontended = true; } boolean collide = false; // True if last slot nonempty for (int attempt = 0; attempt < ATTEMPTS; attempt++) { Buffer<E>[] buffers; Buffer<E> buffer; int n; if (((buffers = table) != null) && ((n = buffers.length) > 0)) { if ((buffer = buffers[(n - 1) & h]) == null) { if ((tableBusy == 0) && casTableBusy()) { // Try to attach new Buffer boolean created = false; try { // Recheck under lock Buffer<E>[] rs; int mask, j; if (((rs = table) != null) && ((mask = rs.length) > 0) && (rs[j = (mask - 1) & h] == null)) { rs[j] = create(e); created = true; } } finally { tableBusy = 0; } if (created) { break; } continue; // Slot is now non-empty } collide = false; } else if (!wasUncontended) { // CAS already known to fail wasUncontended = true; // Continue after rehash } else if (buffer.offer(e) != Buffer.FAILED) { break; } else if (n >= MAXIMUM_TABLE_SIZE || table != buffers) { collide = false; // At max size or stale } else if (!collide) { collide = true; } else if (tableBusy == 0 && casTableBusy()) { try { if (table == buffers) { // Expand table unless stale table = Arrays.copyOf(buffers, n << 1); } } finally { tableBusy = 0; } collide = false; continue; // Retry with expanded table } h = advanceProbe(h); } else if ((tableBusy == 0) && (table == buffers) && casTableBusy()) { boolean init = false; try { // Initialize table if (table == buffers) { @SuppressWarnings({"unchecked", "rawtypes"}) Buffer<E>[] rs = new Buffer[1]; rs[0] = create(e); table = rs; init = true; } } finally { tableBusy = 0; } if (init) { break; } } } }
最后看看RingBuffer
,注意RingBuffer是BoundedBuffer
的内部类。
/** The maximum number of elements per buffer. */ static final int BUFFER_SIZE = 16; // Assume 4-byte references and 64-byte cache line (16 elements per line) //256长度,但是是以16为单位,所以最多存放16个元素 static final int SPACED_SIZE = BUFFER_SIZE << 4; static final int SPACED_MASK = SPACED_SIZE - 1; static final int OFFSET = 16; //RingBuffer数组 final AtomicReferenceArray<E> buffer; //插入方法 @Override public int offer(E e) { long head = readCounter; long tail = relaxedWriteCounter(); //用head和tail来限制个数 long size = (tail - head); if (size >= SPACED_SIZE) { return Buffer.FULL; } //tail追加16 if (casWriteCounter(tail, tail + OFFSET)) { //用tail“取余”得到下标 int index = (int) (tail & SPACED_MASK); //用unsafe.putOrderedObject设值 buffer.lazySet(index, e); return Buffer.SUCCESS; } //如果CAS失败则返回失败 return Buffer.FAILED; } //用consumer来处理buffer的数据 @Override public void drainTo(Consumer<E> consumer) { long head = readCounter; long tail = relaxedWriteCounter(); //判断数据多少 long size = (tail - head); if (size == 0) { return; } do { int index = (int) (head & SPACED_MASK); E e = buffer.get(index); if (e == null) { // not published yet break; } buffer.lazySet(index, null); consumer.accept(e); //head也跟tail一样,每次递增16 head += OFFSET; } while (head != tail); lazySetReadCounter(head); }
注意,ring buffer的size(固定是16个)是不变的,变的是head和tail而已。
总的来说ReadBuffer有如下特点:
- 使用 Striped-RingBuffer来提升对buffer的读写
- 用thread的hash来避开热点key的竞争
- 允许写入的丢失
WriteBuffer
writeBuffer跟readBuffer不一样,主要体现在使用场景的不一样。本来缓存的一般场景是读多写少的,读的并发会更高,且afterRead显得没那么重要,允许延迟甚至丢失。写不一样,写afterWrite不允许丢失,且要求尽量马上执行。Caffeine使用MPSC(Multiple
Producer / Single Consumer)作为buffer数组,实现在MpscGrowableArrayQueue
类,它是仿照JCTools的MpscGrowableArrayQueue来写的。
MPSC允许无锁的高并发写入,但只允许一个消费者,同时也牺牲了部分操作。
MPSC我打算另外分析,这里不展开了。
TimerWheel
除了支持expireAfterAccess
和expireAfterWrite
之外(Guava Cache也支持这两个特性),Caffeine还支持expireAfter
。因为expireAfterAccess和expireAfterWrite都只能是固定的过期时间,这可能满足不了某些场景,譬如记录的过期时间是需要根据某些条件而不一样的,这就需要用户自定义过期时间。
先看看expireAfter的用法
private static LoadingCache<String, String> cache = Caffeine.newBuilder() .maximumSize(256L) .initialCapacity(1) //.expireAfterAccess(2, TimeUnit.DAYS) //.expireAfterWrite(2, TimeUnit.HOURS) .refreshAfterWrite(1, TimeUnit.HOURS) //自定义过期时间 .expireAfter(new Expiry<String, String>() { //返回创建后的过期时间 @Override public long expireAfterCreate(@NonNull String key, @NonNull String value, long currentTime) { return 0; } //返回更新后的过期时间 @Override public long expireAfterUpdate(@NonNull String key, @NonNull String value, long currentTime, @NonNegative long currentDuration) { return 0; } //返回读取后的过期时间 @Override public long expireAfterRead(@NonNull String key, @NonNull String value, long currentTime, @NonNegative long currentDuration) { return 0; } }) .recordStats() .build(new CacheLoader<String, String>() { @Nullable @Override public String load(@NonNull String key) throws Exception { return "value_" + key; } });
通过自定义过期时间,使得不同的key可以动态的得到不同的过期时间。
注意,我把expireAfterAccess和expireAfterWrite注释了,因为这两个特性不能跟expireAfter一起使用。
而当使用了expireAfter特性后,Caffeine会启用一种叫“时间轮”的算法来实现这个功能。更多关于时间轮的介绍,可以看我的文章HashedWheelTimer时间轮原理分析。
好,重点来了,为什么要用时间轮?
对expireAfterAccess和expireAfterWrite的实现是用一个AccessOrderDeque
双端队列,它是FIFO的,因为它们的过期时间是固定的,所以在队列头的数据肯定是最早过期的,要处理过期数据时,只需要首先看看头部是否过期,然后再挨个检查就可以了。但是,如果过期时间不一样的话,这需要对accessOrderQueue进行排序&插入,这个代价太大了。于是,Caffeine用了一种更加高效、优雅的算法-时间轮。
时间轮的结构:
因为在我的对时间轮分析的文章里已经说了时间轮的原理和机制了,所以我就不展开Caffeine对时间轮的实现了。
Caffeine对时间轮的实现在TimerWheel
,它是一种多层时间轮(hierarchical timing wheels )。
看看元素加入到时间轮的schedule
方法:
/** * Schedules a timer event for the node. * * @param node the entry in the cache */ public void schedule(@NonNull Node<K, V> node) { Node<K, V> sentinel = findBucket(node.getVariableTime()); link(sentinel, node); } /** * Determines the bucket that the timer event should be added to. * * @param time the time when the event fires * @return the sentinel at the head of the bucket */ Node<K, V> findBucket(long time) { long duration = time - nanos; int length = wheel.length - 1; for (int i = 0; i < length; i++) { if (duration < SPANS[i + 1]) { long ticks = (time >>> SHIFT[i]); int index = (int) (ticks & (wheel[i].length - 1)); return wheel[i][index]; } } return wheel[length][0]; } /** Adds the entry at the tail of the bucket's list. */ void link(Node<K, V> sentinel, Node<K, V> node) { node.setPreviousInVariableOrder(sentinel.getPreviousInVariableOrder()); node.setNextInVariableOrder(sentinel); sentinel.getPreviousInVariableOrder().setNextInVariableOrder(node); sentinel.setPreviousInVariableOrder(node); }
其他
Caffeine还有其他的优化性能的手段,如使用软引用和弱引用、消除伪共享、CompletableFuture异步等等。
总结
Caffeien是一个优秀的本地缓存,通过使用W-TinyLFU算法, 高性能的readBuffer和WriteBuffer,时间轮算法等,使得它拥有高性能,高命中率(near optimal),低内存占用等特点。
参考资料
TinyLFU论文
Design Of A Modern Cache
Design Of A Modern Cache—Part Deux
Caffeine的github
原作者:AlbenWong
原文链接:Caffeine高性能设计剖析
原出处:个人博客
侵删