1 Overview
Open addressing和Chaining是两种不同的解决hash冲突的策略。当多个不同的key被映射到相同的slot时,chaining方式采用链表保存所有的value。而Open addressing则尝试在该slot的邻近位置查找,直到找到对应的value或者空闲的slot, 这个过程被称作probing。常见的probing策略有Linear probing,Quadratic probing和Double hashing。
2 Chaining
2.1 Chaining in java.util.HashMap
在分析open addressing策略之前,首先简单介绍一下大多数的Java 核心集合类采用的chaining策略,以便比较。 java.util.HashMap有一个Entry[]类型的成员变量table,其每个非null元素都是一个单向链表的表头。
Chaining策略的主要缺点是需要通过Entry保存key,value以及指向链表下个节点的引用(Map.Entry就有四个成员变量),这意味着更多的内存使用(尤其是当key,value本身使用的内存很小时,额外使用的内存所占的比例就显得比较大)。此外链表对CPU的高速缓存不太友好。
3 Open Addressing
3.1 Probing
3.1.1 Linear probing
两次查找位置的间隔为一固定值,即每次查找时在原slot位置的基础上增加一个固定值(通常为1),例如:P = (P + 1) mod SLOT_LENGTH。其最大的优点在于计算速度快,另外对CPU高速缓存更友好。其缺点也非常明显:
假设key1,key2,key3的hash code都相同并且key1被映射到slot(p),那么在计算key2的映射位置时需要查找slot(p), slot(p+1),计算key3的映射位置时需要查找slot(p), slot(p+1),slot(p+2)。也就是说对于导致hash冲突的所有key,在probing过程中会重复查找以前已经查找过的位置,这种现象被称为clustering。
3.1.2 Quadratic probing
两次查找位置的间隔线性增长,例如P(i) = (P + c1*i + c2*i*i) mod SLOT_LENGTH,其中c1和c2为常量且c2不为0(如果为0,那么降级为Linear probing)。 Quadratic probing的各方面性能介于Linear probing和Double hashing之间。
3.1.3 Double hashing
两次查找位置的间隔为一固定值,但是该值通过另外一个hash算法生成,例如P = (P + INCREMENT(key)) mod SLOT_LENGTH,其中INCREMENT即另外一个hash算法。以下是个简单的例子:
H(key) = key mod 10
INCREMENT(key) = 1 + (key mod 7)
P(15): H(15) = 5;
P(35): H(35) = 5, 与P(15)冲突,因此需要进行probe,位置是 (5 + INCREMENT(35)) mod 10 = 6
P(25): H(25) = 5, 与P(15)冲突,因此需要进行probe,位置是 (5 + INCREMENT(25)) mod 10 = 0
P(75): H(75) = 5, 与P(15)冲突,因此需要进行probe,位置是 (5 + INCREMENT(75)) mod 10 = 1
从以上例子可以看出,跟Linear probing相比,减少了重复查找的次数。
3.2 Load Factor
基于open addressing的哈希表的性能对其load factor属性值非常敏感。如果该值超过0.7 (Trove maps/sets的默认load factor是0.5),那么性能会下降的非常明显。由于hash冲突导致的probing次数跟(loadFactor) / (1 - loadFactor)成正比。当loadFactor为1时,如果哈希表中的空闲slot非常少,那么可能会导致probing的次数非常大。
3.3 Open addressing in gnu.trove.THashMap
GNU Trove (http://trove4j.sourceforge.net/) 是一个Java 集合类库。在某些场景下,Trove集合类库提供了更好的性能,而且内存使用更少。以下是Trove中跟open addressing相关的几个特性:
TObjectHashingStrategy
接口, Trove支持定制hash算法(例如不希望使用String或者数组的默认hash算法)。跟java.util.HashMap相比,gnu.trove.THashMap没有Entry[] table之类的成员变量,而是分别通过Object[] _set,V[] _values直接保存key和value。在逻辑上,Object[] _set中的每个元素都有三种状态:
这三种状态的迁移过程如下:
以下是关于状态迁移的简单例子(:= 的含义是赋值, H(key) = key mod 11):
以下是与get()方法相关的代码片段:
public V get(Object key) { int index = index((K) key); return index < 0 ? null : _values[index]; } protected int index(T obj) { final TObjectHashingStrategy<T> hashing_strategy = _hashingStrategy; final Object[] set = _set; final int length = set.length; final int hash = hashing_strategy.computeHashCode(obj) & 0x7fffffff; int index = hash % length; Object cur = set[index]; if ( cur == FREE ) return -1; // NOTE: here it has to be REMOVED or FULL (some user-given value) if ( cur == REMOVED || ! hashing_strategy.equals((T) cur, obj)) { // see Knuth, p. 529 final int probe = 1 + (hash % (length - 2)); do { index -= probe; if (index < 0) { index += length; } cur = set[index]; } while (cur != FREE && (cur == REMOVED || ! _hashingStrategy.equals((T) cur, obj))); } return cur == FREE ? -1 : index; }
从以上代码可以看出get()方法的流程如下, 根据key的hash值找到对应的set元素,判断是否存在hash冲突。
如果不存在hash冲突,那么该set元素的可能状态如下:
以下是与put()方法相关的代码片段:
public V put(K key, V value) { int index = insertionIndex(key); return doPut(key, value, index); } private V doPut(K key, V value, int index) { V previous = null; Object oldKey; boolean isNewMapping = true; if (index < 0) { index = -index -1; previous = _values[index]; isNewMapping = false; } oldKey = _set[index]; _set[index] = key; _values[index] = value; if (isNewMapping) { postInsertHook(oldKey == FREE); } return previous; } protected int insertionIndex(T obj) { final TObjectHashingStrategy<T> hashing_strategy = _hashingStrategy; final Object[] set = _set; final int length = set.length; final int hash = hashing_strategy.computeHashCode(obj) & 0x7fffffff; int index = hash % length; Object cur = set[index]; if (cur == FREE) { return index; // empty, all done } else if (cur != REMOVED && hashing_strategy.equals((T) cur, obj)) { return -index -1; // already stored } else { // already FULL or REMOVED, must probe // compute the double hash final int probe = 1 + (hash % (length - 2)); // if the slot we landed on is FULL (but not removed), probe // until we find an empty slot, a REMOVED slot, or an element // equal to the one we are trying to insert. // finding an empty slot means that the value is not present // and that we should use that slot as the insertion point; // finding a REMOVED slot means that we need to keep searching, // however we want to remember the offset of that REMOVED slot // so we can reuse it in case a "new" insertion (i.e. not an update) // is possible. // finding a matching value means that we've found that our desired // key is already in the table if (cur != REMOVED) { // starting at the natural offset, probe until we find an // offset that isn't full. do { index -= probe; if (index < 0) { index += length; } cur = set[index]; } while (cur != FREE && cur != REMOVED && ! hashing_strategy.equals((T) cur, obj)); } // if the index we found was removed: continue probing until we // locate a free location or an element which equal()s the // one we have. if (cur == REMOVED) { int firstRemoved = index; while (cur != FREE && (cur == REMOVED || ! hashing_strategy.equals((T) cur, obj))) { index -= probe; if (index < 0) { index += length; } cur = set[index]; } // NOTE: cur cannot == REMOVED in this block return (cur != FREE) ? -index -1 : firstRemoved; } // if it's full, the key is already stored // NOTE: cur cannot equal REMOVE here (would have retuned already (see above) return (cur != FREE) ? -index -1 : index; } }
从以上代码可以看出,THashMap使用Double hashing。用来计算增量的hash算法是final int probe = 1 + (hash % (length - 2)); 如果insertionIndex()方法的返回值为正值,那么该值就是可用的slot位置;如果为负值,那么说明该key之前已经保存过,(-index-1)就是之前的slot位置。
put()方法的流程如下, 根据key的hash值找到对应的set元素,判断是否存在hash冲突。
如果不存在hash冲突,那么该set元素的可能状态如下: