一个内存池的实现实例

内部构造

内存池类MemoryPool的声明如下：

class MemoryPool{private: MemoryBlock* pBlock; USHORT nUnitSize; USHORT nInitSize; USHORT nGrowSize;public: MemoryPool( USHORT nUnitSize, USHORT nInitSize = 1024, USHORT nGrowSize = 256 ); ~MemoryPool(); void* Alloc(); void Free( void* p );};

MemoryBlock为内存池中附着在真正用来为内存请求分配内存的内存块头部的结构体，它描述了与之联系的内存块的使用信息：

struct MemoryBlock{ USHORT nSize; USHORT nFree; USHORT nFirst; USHORT nDummyAlign1; MemoryBlock* pNext; char aData[1]; static void* operator new(size_t, USHORT nTypes, USHORT nUnitSize) { return ::operator new(sizeof(MemoryBlock) + nTypes * nUnitSize); } static void operator delete(void *p, size_t) { ::operator delete (p); } MemoryBlock (USHORT nTypes = 1, USHORT nUnitSize = 0); ~MemoryBlock() {}};

此内存池的数据结构如图所示。

 内存池的数据结构

 总体机制

此内存池的总体机制如下。

（1）在运行过程中，MemoryPool内存池可能会有多个用来满足内存申请请求的内存块，这些内存块是从进程堆中开辟的一个较大的连续内存区域，它由一个MemoryBlock结构体和多个可供分配的内存单元组成，所有内存块组成了一个内存块链表，MemoryPool的pBlock是这个链表的头。对每个内存块，都可以通过其头部的MemoryBlock结构体的pNext成员访问紧跟在其后面的那个内存块。

（2）每个内存块由两部分组成，即一个MemoryBlock结构体和多个内存分配单元。这些内存分配单元大小固定（由MemoryPool的nUnitSize表示），MemoryBlock结构体并不维护那些已经分配的单元的信息；相反，它只维护没有分配的自由分配单元的信息。它有两个成员比较重要：nFree和nFirst。nFree记录这个内存块中还有多少个自由分配单元，而nFirst则记录下一个可供分配的单元的编号。每一个自由分配单元的头两个字节（即一个USHORT型值）记录了紧跟它之后的下一个自由分配单元的编号，这样，通过利用每个自由分配单元的头两个字节，一个MemoryBlock中的所有自由分配单元被链接起来。

（3）当有新的内存请求到来时，MemoryPool会通过pBlock遍历MemoryBlock链表，直到找到某个MemoryBlock所在的内存块，其中还有自由分配单元（通过检测MemoryBlock结构体的nFree成员是否大于0）。如果找到这样的内存块，取得其MemoryBlock的nFirst值（此为该内存块中第1个可供分配的自由单元的编号）。然后根据这个编号定位到该自由分配单元的起始位置（因为所有分配单元大小固定，因此每个分配单元的起始位置都可以通过编号分配单元大小来偏移定位），这个位置就是用来满足此次内存申请请求的内存的起始地址。但在返回这个地址前，需要首先将该位置开始的头两个字节的值（这两个字节值记录其之后的下一个自由分配单元的编号）赋给本内存块的MemoryBlock的nFirst成员。这样下一次的请求就会用这个编号对应的内存单元来满足，同时将此内存块的MemoryBlock的nFree递减1，然后才将刚才定位到的内存单元的起始位置作为此次内存请求的返回地址返回给调用者。

（4）如果从现有的内存块中找不到一个自由的内存分配单元（当第1次请求内存，以及现有的所有内存块中的所有内存分配单元都已经被分配时会发生这种情形），MemoryPool就会从进程堆中申请一个内存块（这个内存块包括一个MemoryBlock结构体，及紧邻其后的多个内存分配单元，假设内存分配单元的个数为n，n可以取值MemoryPool中的nInitSize或者nGrowSize），申请完后，并不会立刻将其中的一个分配单元分配出去，而是需要首先初始化这个内存块。初始化的操作包括设置MemoryBlock的nSize为所有内存分配单元的大小（注意，并不包括MemoryBlock结构体的大小）、nFree为n-1（注意，这里是n-1而不是n，因为此次新内存块就是为了满足一次新的内存请求而申请的，马上就会分配一块自由存储单元出去，如果设为n-1，分配一个自由存储单元后无须再将n递减1），nFirst为1（已经知道nFirst为下一个可以分配的自由存储单元的编号。为1的原因与nFree为n-1相同，即立即会将编号为0的自由分配单元分配出去。现在设为1，其后不用修改nFirst的值），MemoryBlock的构造需要做更重要的事情，即将编号为0的分配单元之后的所有自由分配单元链接起来。如前所述，每个自由分配单元的头两个字节用来存储下一个自由分配单元的编号。另外，因为每个分配单元大小固定，所以可以通过其编号和单元大小（MemoryPool的nUnitSize成员）的乘积作为偏移值进行定位。现在唯一的问题是定位从哪个地址开始？答案是MemoryBlock的aData[1]成员开始。因为aData[1]实际上是属于MemoryBlock结构体的（MemoryBlock结构体的最后一个字节），所以实质上，MemoryBlock结构体的最后一个字节也用做被分配出去的分配单元的一部分。因为整个内存块由MemoryBlock结构体和整数个分配单元组成，这意味着内存块的最后一个字节会被浪费，这个字节在图中用位于两个内存的最后部分的浓黑背景的小块标识。确定了分配单元的起始位置后，将自由分配单元链接起来的工作就很容易了。即从aData位置开始，每隔nUnitSize大小取其头两个字节，记录其之后的自由分配单元的编号。因为刚开始所有分配单元都是自由的，所以这个编号就是自身编号加1，即位置上紧跟其后的单元的编号。初始化后，将此内存块的第1个分配单元的起始地址返回，已经知道这个地址就是aData。

（5）当某个被分配的单元因为delete需要回收时，该单元并不会返回给进程堆，而是返回给MemoryPool。返回时，MemoryPool能够知道该单元的起始地址。这时，MemoryPool开始遍历其所维护的内存块链表，判断该单元的起始地址是否落在某个内存块的地址范围内。如果不在所有内存地址范围内，则这个被回收的单元不属于这个MemoryPool；如果在某个内存块的地址范围内，那么它会将这个刚刚回收的分配单元加到这个内存块的MemoryBlock所维护的自由分配单元链表的头部，同时将其nFree值递增1。回收后，考虑到资源的有效利用及后续操作的性能，内存池的操作会继续判断：如果此内存块的所有分配单元都是自由的，那么这个内存块就会从MemoryPool中被移出并作为一个整体返回给进程堆；如果该内存块中还有非自由分配单元，这时不能将此内存块返回给进程堆。但是因为刚刚有一个分配单元返回给了这个内存块，即这个内存块有自由分配单元可供下次分配，因此它会被移到MemoryPool维护的内存块的头部。这样下次的内存请求到来，MemoryPool遍历其内存块链表以寻找自由分配单元时，第1次寻找就会找到这个内存块。因为这个内存块确实有自由分配单元，这样可以减少MemoryPool的遍历次数。

综上所述，每个内存池（MemoryPool）维护一个内存块链表（单链表），每个内存块由一个维护该内存块信息的块头结构（MemoryBlock）和多个分配单元组成，块头结构MemoryBlock则进一步维护一个该内存块的所有自由分配单元组成的"链表"。这个链表不是通过"指向下一个自由分配单元的指针"链接起来的，而是通过"下一个自由分配单元的编号"链接起来，这个编号值存储在该自由分配单元的头两个字节中。另外，第1个自由分配单元的起始位置并不是MemoryBlock结构体"后面的"第1个地址位置，而是MemoryBlock结构体"内部"的最后一个字节aData（也可能不是最后一个，因为考虑到字节对齐的问题），即分配单元实际上往前面错了一位。又因为MemoryBlock结构体后面的空间刚好是分配单元的整数倍，这样依次错位下去，内存块的最后一个字节实际没有被利用。这么做的一个原因也是考虑到不同平台的移植问题，因为不同平台的对齐方式可能不尽相同。即当申请MemoryBlock大小内存时，可能会返回比其所有成员大小总和还要大一些的内存。最后的几个字节是为了"补齐"，而使得aData成为第1个分配单元的起始位置，这样在对齐方式不同的各种平台上都可以工作。

 细节剖析

有了上述的总体印象后，仔细剖析其实现细节。

（1）MemoryPool的构造如下：

MemoryPool::MemoryPool( USHORT _nUnitSize, USHORT _nInitSize, USHORT _nGrowSize ){ pBlock = NULL; ① nInitSize = _nInitSize; ② nGrowSize = _nGrowSize; ③ if ( _nUnitSize > 4 ) nUnitSize = (_nUnitSize + (MEMPOOL_ALIGNMENT-1)) & ~(MEMPOOL_ALIGNMENT-1); ④ else if ( _nUnitSize <= 2 ) nUnitSize = 2; ⑤ else nUnitSize = 4;}

从①处可以看出，MemoryPool创建时，并没有立刻创建真正用来满足内存申请的内存块，即内存块链表刚开始时为空。

②处和③处分别设置"第1次创建的内存块所包含的分配单元的个数"，及"随后创建的内存块所包含的分配单元的个数"，这两个值在MemoryPool创建时通过参数指定，其后在该MemoryPool对象生命周期中一直不变。

后面的代码用来设置nUnitSize，这个值参考传入的_nUnitSize参数。但是还需要考虑两个因素。如前所述，每个分配单元在自由状态时，其头两个字节用来存放"其下一个自由分配单元的编号"。即每个分配单元"最少"有"两个字节"，这就是⑤处赋值的原因。④处是将大于4个字节的大小_nUnitSize往上"取整到"大于_nUnitSize的最小的MEMPOOL_ ALIGNMENT的倍数（前提是MEMPOOL_ALIGNMENT为2的倍数）。如_nUnitSize为11时，MEMPOOL_ALIGNMENT为8，nUnitSize为16；MEMPOOL_ALIGNMENT为4，nUnitSize为12；MEMPOOL_ALIGNMENT为2，nUnitSize为12，依次类推。

（2）当向MemoryPool提出内存请求时：

void* MemoryPool::Alloc(){ if ( !pBlock ) ① { …… } MemoryBlock* pMyBlock = pBlock; while (pMyBlock && !pMyBlock->nFree )② pMyBlock = pMyBlock->pNext; if ( pMyBlock ) ③ { char* pFree = pMyBlock->aData+(pMyBlock->nFirst*nUnitSize); pMyBlock->nFirst = *((USHORT*)pFree); pMyBlock->nFree--; return (void*)pFree; } else ④ { if ( !nGrowSize ) return NULL; pMyBlock = new(nGrowSize, nUnitSize) FixedMemBlock(nGrowSize, nUnitSize); if ( !pMyBlock ) return NULL; pMyBlock->pNext = pBlock; pBlock = pMyBlock; return (void*)(pMyBlock->aData); }}

MemoryPool满足内存请求的步骤主要由四步组成。

①处首先判断内存池当前内存块链表是否为空，如果为空，则意味着这是第1次内存申请请求。这时，从进程堆中申请一个分配单元个数为nInitSize的内存块，并初始化该内存块（主要初始化MemoryBlock结构体成员，以及创建初始的自由分配单元链表，下面会详细分析其代码）。如果该内存块申请成功，并初始化完毕，返回第1个分配单元给调用函数。第1个分配单元以MemoryBlock结构体内的最后一个字节为起始地址。

②处的作用是当内存池中已有内存块（即内存块链表不为空）时遍历该内存块链表，寻找还有"自由分配单元"的内存块。

③处检查如果找到还有自由分配单元的内存块，则"定位"到该内存块现在可以用的自由分配单元处。"定位"以MemoryBlock结构体内的最后一个字节位置aData为起始位置，以MemoryPool的nUnitSize为步长来进行。找到后，需要修改MemoryBlock的nFree信息（剩下来的自由分配单元比原来减少了一个），以及修改此内存块的自由存储单元链表的信息。在找到的内存块中，pMyBlock->nFirst为该内存块中自由存储单元链表的表头，其下一个自由存储单元的编号存放在pMyBlock->nFirst指示的自由存储单元（亦即刚才定位到的自由存储单元）的头两个字节。通过刚才定位到的位置，取其头两个字节的值，赋给pMyBlock->nFirst，这就是此内存块的自由存储单元链表的新的表头，即下一次分配出去的自由分配单元的编号（如果nFree大于零的话）。修改维护信息后，就可以将刚才定位到的自由分配单元的地址返回给此次申请的调用函数。注意，因为这个分配单元已经被分配，而内存块无须维护已分配的分配单元，因此该分配单元的头两个字节的信息已经没有用处。换个角度看，这个自由分配单元返回给调用函数后，调用函数如何处置这块内存，内存池无从知晓，也无须知晓。此分配单元在返回给调用函数时，其内容对于调用函数来说是无意义的。因此几乎可以肯定调用函数在用这个单元的内存时会覆盖其原来的内容，即头两个字节的内容也会被抹去。因此每个存储单元并没有因为需要链接而引入多余的维护信息，而是直接利用单元内的头两个字节，当其分配后，头两个字节也可以被调用函数利用。而在自由状态时，则用来存放维护信息，即下一个自由分配单元的编号，这是一个有效利用内存的好例子。

④处表示在②处遍历时，没有找到还有自由分配单元的内存块，这时，需要重新向进程堆申请一个内存块。因为不是第一次申请内存块，所以申请的内存块包含的分配单元个数为nGrowSize，而不再是nInitSize。与①处相同，先做这个新申请内存块的初始化工作，然后将此内存块插入MemoryPool的内存块链表的头部，再将此内存块的第1个分配单元返回给调用函数。将此新内存块插入内存块链表的头部的原因是该内存块还有很多可供分配的自由分配单元（除非nGrowSize等于1，这应该不太可能。因为内存池的含义就是一次性地从进程堆中申请一大块内存，以供后续的多次申请），放在头部可以使得在下次收到内存申请时，减少②处对内存块的遍历时间。

可以MemoryPool来展示MemoryPool::Alloc的过程。某个时刻MemoryPool的内部状态。

某个时刻MemoryPool的内部状态

因为MemoryPool的内存块链表不为空，因此会遍历其内存块链表。又因为第1个内存块里有自由的分配单元，所以会从第1个内存块中分配。检查nFirst，其值为m，这时pBlock->aData+(pBlock->nFirst*nUnitSize)定位到编号为m的自由分配单元的起始位置（用pFree表示）。在返回pFree之前，需要修改此内存块的维护信息。首先将nFree递减1，然后取得pFree处开始的头两个字节的值（需要说明的是，这里aData处值为k。其实不是这一个字节。而是以aData和紧跟其后的另外一个字节合在一起构成的一个USHORT的值，不可误会）。发现为k，这时修改pBlock的nFirst为k。然后，返回pFree。此时MemoryPool的结构如图所示。

 MemoryPool的结构

可以看到，原来的第1个可供分配的单元（m编号处）已经显示为被分配的状态。而pBlock的nFirst已经指向原来m单元下一个自由分配单元的编号，即k。

（3）MemoryPool回收内存时：

void MemoryPool::Free( void* pFree ){ …… MemoryBlock* pMyBlock = pBlock; while ( ((ULONG)pMyBlock->aData > (ULONG)pFree) || ((ULONG)pFree >= ((ULONG)pMyBlock->aData + pMyBlock->nSize)) )① { …… } pMyBlock->nFree++; ② *((USHORT*)pFree) = pMyBlock->nFirst; ③ pMyBlock->nFirst = (USHORT)(((ULONG)pFree-(ULONG)(pBlock->aData)) / nUnitSize);④ if (pMyBlock->nFree*nUnitSize == pMyBlock->nSize )⑤ { …… } else { …… }}

如前所述，回收分配单元时，可能会将整个内存块返回给进程堆，也可能将被回收分配单元所属的内存块移至内存池的内存块链表的头部。这两个操作都需要修改链表结构。这时需要知道该内存块在链表中前一个位置的内存块。

①处遍历内存池的内存块链表，确定该待回收分配单元（pFree）落在哪一个内存块的指针范围内，通过比较指针值来确定。

运行到②处，pMyBlock即找到的包含pFree所指向的待回收分配单元的内存块（当然，这时应该还需要检查pMyBlock为NULL时的情形，即pFree不属于此内存池的范围，因此不能返回给此内存池，读者可以自行加上）。这时将pMyBlock的nFree递增1，表示此内存块的自由分配单元多了一个。

③处用来修改该内存块的自由分配单元链表的信息，它将这个待回收分配单元的头两个字节的值指向该内存块原来的第一个可分配的自由分配单元的编号。

④处将pMyBlock的nFirst值改变为指向这个待回收分配单元的编号，其编号通过计算此单元的起始位置相对pMyBlock的aData位置的差值，然后除以步长（nUnitSize）得到。

实质上，③和④两步的作用就是将此待回收分配单元"真正回收"。值得注意的是，这两步实际上是使得此回收单元成为此内存块的下一个可分配的自由分配单元，即将它放在了自由分配单元链表的头部。注意，其内存地址并没有发生改变。实际上，一个分配单元的内存地址无论是在分配后，还是处于自由状态时，一直都不会变化。变化的只是其状态（已分配/自由），以及当其处于自由状态时在自由分配单元链表中的位置。

⑤处检查当回收完毕后，包含此回收单元的内存块的所有单元是否都处于自由状态，且此内存是否处于内存块链表的头部。如果是，将此内存块整个的返回给进程堆，同时修改内存块链表结构。

注意，这里在判断一个内存块的所有单元是否都处于自由状态时，并没有遍历其所有单元，而是判断nFree乘以nUnitSize是否等于nSize。nSize是内存块中所有分配单元的大小，而不包括头部MemoryBlock结构体的大小。这里可以看到其用意，即用来快速检查某个内存块中所有分配单元是否全部处于自由状态。因为只需结合nFree和nUnitSize来计算得出结论，而无须遍历和计算所有自由状态的分配单元的个数。

另外还需注意的是，这里并不能比较nFree与nInitSize或nGrowSize的大小来判断某个内存块中所有分配单元都为自由状态，这是因为第1次分配的内存块（分配单元个数为nInitSize）可能被移到链表的后面，甚至可能在移到链表后面后，因为某个时间其所有单元都处于自由状态而被整个返回给进程堆。即在回收分配单元时，无法判定某个内存块中的分配单元个数到底是nInitSize还是nGrowSize，也就无法通过比较nFree与nInitSize或nGrowSize的大小来判断一个内存块的所有分配单元是否都为自由状态。

以上面分配后的内存池状态作为例子，假设这时第2个内存块中的最后一个单元需要回收（已被分配，假设其编号为m，pFree指针指向它），如图所示。

不难发现，这时nFirst的值由原来的0变为m。即此内存块下一个被分配的单元是m编号的单元，而不是0编号的单元（最先分配的是最新回收的单元，从这一点看，这个过程与栈的原理类似，即先进后出。只不过这里的"进"意味着"回收"，而"出"则意味着"分配"）。相应地，m的"下一个自由单元"标记为0，即内存块原来的"下一个将被分配出去的单元"，这也表明最近回收的分配单元被插到了内存块的"自由分配单元链表"的头部。当然，nFree递增1。

分配后的内存池状态

处理至⑥处之前，其状态如图所示。

图 处理至⑥处之前的内存池状态

这里需要注意的是，虽然pFree被"回收"，但是pFree仍然指向m编号的单元，这个单元在回收过程中，其头两个字节被覆写，但其他部分的内容并没有改变。而且从整个进程的内存使用角度来看，这个m编号的单元的状态仍然是"有效的"。因为这里的"回收"只是回收给了内存池，而并没有回收给进程堆，因此程序仍然可以通过pFree访问此单元。但是这是一个很危险的操作，因为首先该单元在回收过程中头两个字节已被覆写，并且该单元可能很快就会被内存池重新分配。因此回收后通过pFree指针对这个单元的访问都是错误的，读操作会读到错误的数据，写操作则可能会破坏程序中其他地方的数据，因此需要格外小心。

接着，需要判断该内存块的内部使用情况，及其在内存块链表中的位置。如果该内存块中省略号"……"所表示的其他部分中还有被分配的单元，即nFree乘以nUnitSize不等于nSize。因为此内存块不在链表头，因此还需要将其移到链表头部，如图所示。

图因回收引起的MemoryBlock移动

如果该内存块中省略号"……"表示的其他部分中全部都是自由分配单元，即nFree乘以nUnitSize等于nSize。因为此内存块不在链表头，所以此时需要将此内存块整个回收给进程堆，回收后内存池的结构如图所示。

图 回收后内存池的结构

一个内存块在申请后会初始化，主要是为了建立最初的自由分配单元链表，下面是其详细代码：

MemoryBlock::MemoryBlock (USHORT nTypes, USHORT nUnitSize) : nSize (nTypes * nUnitSize), nFree (nTypes - 1), ④ nFirst (1), ⑤ pNext (0){ char * pData = aData; ① for (USHORT i = 1; i < nTypes; i++) ② { *reinterpret_cast<USHORT*>(pData) = i; ③ pData += nUnitSize; }}

这里可以看到，①处pData的初值是aData，即0编号单元。但是②处的循环中i却是从1开始，然后在循环内部的③处将pData的头两个字节值置为i。即0号单元的头两个字节值为1，1号单元的头两个字节值为2，一直到（nTypes-2）号单元的头两个字节值为（nTypes-1）。这意味着内存块初始时，其自由分配单元链表是从0号开始。依次串联，一直到倒数第2个单元指向最后一个单元。

还需要注意的是，在其初始化列表中，nFree初始化为nTypes-1（而不是nTypes），nFirst初始化为1（而不是0）。这是因为第1个单元，即0编号单元构造完毕后，立刻会被分配。另外注意到最后一个单元初始并没有设置头两个字节的值，因为该单元初始在本内存块中并没有下一个自由分配单元。但是从上面例子中可以看到，当最后一个单元被分配并回收后，其头两个字节会被设置。

图所示为一个内存块初始化后的状态。

图 一个内存块初始化后的状态

当内存池析构时，需要将内存池的所有内存块返回给进程堆：

MemoryPool::~MemoryPool(){ MemoryBlock* pMyBlock = pBlock; while ( pMyBlock ) { …… }}

 使用方法

从上面的分析可以看到，该内存池主要有两个对外接口函数，即Alloc和Free。Alloc返回所申请的分配单元（固定大小内存），Free则回收传入的指针代表的分配单元的内存给内存池。分配的信息则通过MemoryPool的构造函数指定，包括分配单元大小、内存池第1次申请的内存块中所含分配单元的个数，以及内存池后续申请的内存块所含分配单元的个数等。

综上所述，当需要提高某些关键类对象的申请／回收效率时，可以考虑将该类所有生成对象所需的空间都从某个这样的内存池中开辟。在销毁对象时，只需要返回给该内存池。"一个类的所有对象都分配在同一个内存池对象中"这一需求很自然的设计方法就是为这样的类声明一个静态内存池对象，同时为了让其所有对象都从这个内存池中开辟内存，而不是缺省的从进程堆中获得，需要为该类重载一个new运算符。因为相应地，回收也是面向内存池，而不是进程的缺省堆，还需要重载一个delete运算符。在new运算符中用内存池的Alloc函数满足所有该类对象的内存请求，而销毁某对象则可以通过在delete运算符中调用内存池的Free完成。

性能比较

为了测试利用内存池后的效果，通过一个很小的测试程序可以发现采用内存池机制后耗时为297 ms。而没有采用内存池机制则耗时625 ms，速度提高了52.48%。速度提高的原因可以归结为几点，其一，除了偶尔的内存申请和销毁会导致从进程堆中分配和销毁内存块外，绝大多数的内存申请和销毁都由内存池在已经申请到的内存块中进行，而没有直接与进程堆打交道，而直接与进程堆打交道是很耗时的操作；其二，这是单线程环境的内存池，可以看到内存池的Alloc和Free操作中并没有加线程保护措施。因此如果类A用到该内存池，则所有类A对象的创建和销毁都必须发生在同一个线程中。但如果类A用到内存池，类B也用到内存池，那么类A的使用线程可以不必与类B的使用线程是同一个线程。

因为内存池技术使得同类型的对象分布在相邻的内存区域，而程序会经常对同一类型的对象进行遍历操作。因此在程序运行过程中发生的缺页应该会相应少一些，但这个一般只能在真实的复杂应用环境中进行验证。内存的申请和释放对一个应用程序的整体性能影响极大，甚至在很多时候成为某个应用程序的瓶颈。消除内存申请和释放引起的瓶颈的方法往往是针对内存使用的实际情况提供一个合适的内存池。内存池之所以能够提高性能，主要是因为它能够利用应用程序的实际内存使用场景中的某些"特性"。比如某些内存申请与释放肯定发生在一个线程中，某种类型的对象生成和销毁与应用程序中的其他类型对象要频繁得多，等等。针对这些特性，可以为这些特殊的内存使用场景提供量身定做的内存池。这样能够消除系统提供的缺省内存机制中，对于该实际应用场景中的不必要的操作，从而提升应用程序的整体性能。

Many enterprise search vendors have announced that clustering of search results is now part of their product and user experience. The most recent case is Google (press center, blog post, blogosphere reaction). Microsoft researchers have also experimented with clustering, without these experiments finding their way yet into Microsoft’s products.

By definition, a clustering engine analyzes the top (say 200-500) search results from a query and displays the main themes, typically as folders that may consist of subfolders.

The spread of clustering engines is gratifying, since Vivisimo was founded on a breakthrough clustering algorithm, has been refining the approach and educating and selling into the search market since 2001, and has evolved into a complete enterprise search provider.

Just as with search results, or as with any other designed product, judging the quality of a clustering engine requires some skill. Before judging quality, let’s first explore clustering’s end user value, that is, how it enhances knowledge worker productivity.

Clustering enhances end user productivity in at least three ways:

1. At a glance, users gain an easy overview of the main distinct themes that are present in the top search results.
2. By clicking on clusters that satisfy their needs or interests, users can quickly arrive at search results that are valuable but low ranked, say, #73 or even #429 in the results list, and so would never be noticed otherwise. The user’s visibility into the content is greatly enhanced.
3. After arriving at a cluster or sub-cluster, related results are placed together (”clustered”) rather than scattered throughout the ranked list. This expedites finding related or the best content.

In short, clustering lets users overview, find, discover, and compare information more productively. How does clustering quality enhance or detract from this productivity gain?

1. To grasp an overview of the main themes, the cluster labels should be concise and natural-looking. Also, the clusters shouldn’t overlap too much in their contents, otherwise the user will be overloaded with too many clusters expressing overly related themes. If the clusters don’t overlap too much, so that on average a search result appears in only 1.2 to 1.5 clusters, then the main distinct themes will be shown, rather than similar/duplicate themes on overlapping content. Also, the cluster labels shouldn’t be artificially limited to labels that contain the query word, or labels that have two or more words in them, or some other artifact of an inferior clustering approach. Finally, the underlying search engine snippets (aka excerpts or dynamic summaries) should be full enough so that clustering has enough input text to work with.
2. To arrive at low-ranked but valuable results, the clustering engine should be fast enough so that 200-500 results or more can be clustered within an acceptable response time. If user authentication is an issue (note discussion here), then the response time should include the time for the search engine to verify that the user can view these documents. Also, the cluster labels should accurately express its contents, otherwise the user wastes time on wild goose chases.
3. In order for similar results to be placed together in the same clusters, the clustering software should possess the linguistic knowledge needed to correctly handle cases like the following:
   • sort out the meanings in middle ages, middle aged, and medieval, and in news release, new release, and press release.
   • realize that king and kingship are very related, unlike gun and gunship.
   • realize that unfearful and fearless are synonymous, but not unhelpful and helpless.
   • plus many thousands of other linguistic relationships that take time and background knowledge to learn, whether by humans in school or by computers.

There are endless other subtleties, but enough: what’s the bottom line? Here are some questions to ask about the quality of a clustering search engine:

1. Are the cluster folders determined by analyzing the top search results? If not, then no overview of the major themes is being given. Instead, the “folders” are probably based on query logs.
2. Does clicking on a cluster cause a new search to be done? If so, it’s not clustering but something else, likely query refinement, which leads to a discontinuous user experience.
3. Is the clustering engine able to handle 200-500 results or even more?
4. Are the cluster names concise and natural-looking, and do they correctly handle numbers, punctuation, diacritics, foreign words, etc.?
5. Is there evidence that the clustering engine possesses considerable linguistic knowledge? And in other languages besides English, if needed? For example, are many (30-40% or more) of the search results left unclustered into an Other category? This suggests a deficiency in detecting related meanings.

Here are some queries to try:

1. bird flu vaccine
2. shi’ite sunni rivalry
3. 9/11 hijackers
4. C++
5. “Raúl González Blanco”
6. bill gates

YouTube uses Apache for FastCGI serving. (I wonder if things would have been easier for them had they chosen nginx, which is apparently wonderful for FastCGI and less problematic than Lighttpd)

YouTube is coded mostly in Python. Why? “Development speed critical”.

They use psyco, Python -> C compiler, and also C extensions, for performance critical work.

They use Lighttpd for serving the video itself, for a big improvement over Apache.

Each video hosted by a “mini cluster”, which is a set of machine with the same content. This is a simple way to provide headroom (slack), so that a machine can be taken down for maintenance (or can fail) without affecting users. It also provides a form of backup.

The most popular videos are on a CDN (Content Distribution Network) - they use external CDNs and well as Google’s CDN. Requests to their own machines are therefore tail-heavy (in the “Long Tail” sense), because the head codes to the CDN instead.

Because of the tail-heavy load, random disks seeks are especially important (perhaps more important than caching?).

YouTube uses simple, cheap, commodity Hardware. The more expensive the hardware, the more expensive everything else gets (support, etc.). Maintenance is mostly done with rsync, SSH, other simple, common tools.
The fun is not over: Cuong showed a recent email titled “3 days of video storage left”. There is constant work to keep up with the growth.

Thumbnails turn out to be surprisingly hard to serve efficiently. Because there, on average, 4 thumbnails per video and many thumbnails per pages, the overall number of thumbnails per second is enormous. They use a separate group of machines to serve thumbnails, with extensive caching and OS tuning specific to this load.

YouTube was bit by a “too many files in one dir” limit: at one point they could accept no more uploads (!!) because of this. The first fix was the usual one: split the files across many directories, and switch to another file system better suited for many small files.

Cuong joked about “The Windows approach of scaling: restart everything”

Lighttpd turned out to be poor for serving the thumbnails, because its main loop is a bottleneck to load files from disk; they addressed this by modifying Lighttpd to add worker threads to read from disk. This was good but still not good enough, with one thumbnail per file, because the enormous number of files was terribly slow to work with (imagine tarring up many million files).

Their new solution for thumbnails is to use Google’s BigTable, which provides high performance for a large number of rows, fault tolerance, caching, etc. This is a nice (and rare?) example of actual synergy in an acquisition.

YouTube uses MySQL to store metadata. Early on they hit a Linux kernel issue which prioritized the page cache higher than app data, it swapped out the app data, totally overwhelming the system. They recovered from this by removing the swap partition (while live!). This worked.

YouTube uses Memcached.

To scale out the database, they first used MySQL replication. Like everyone else that goes down this path, they eventually reach a point where replicating the writes to all the DBs, uses up all the capacity of the slaves. They also hit a issue with threading and replication, which they worked around with a very clever “cache primer thread” working a second or so ahead of the replication thread, prefetching the data it would need.

As the replicate-one-DB approach faltered, they resorted to various desperate measures, such as splitting the video watching in to a separate set of replicas, intentionally allowing the non-video-serving parts of YouTube to perform badly so as to focus on serving videos.

Their initial MySQL DB server configuration had 10 disks in a RAID10. This does not work very well, because the DB/OS can’t take advantage of the multiple disks in parallel. They moved to a set of RAID1s, appended together. In my experience, this is better, but still not great. An approach that usually works even better is to intentionally split different data on to different RAIDs: for example, a RAID for the OS / application, a RAID for the DB logs, one or more RAIDs for the DB table (uses “tablespaces” to get your #1 busiest table on separate spindles from your #2 busiest table), one or more RAID for index, etc. Big-iron Oracle installation sometimes take this approach to extremes; the same thing can be done with free DBs on free OSs also.

In spite of all these effort, they reached a point where replication of one large DB was no longer able to keep up. Like everyone else, they figured out that the solution database partitioning in to “shards”. This spread reads and writes in to many different databases (on different servers) that are not all running each other’s writes. The result is a large performance boost, better cache locality, etc. YouTube reduced their total DB hardware by 30% in the process.

It is important to divide users across shards by a controllable lookup mechanism, not only by a hash of the username/ID/whatever, so that you can rebalance shards incrementally.

An interesting DMCA issue: YouTube complies with takedown requests; but sometimes the videos are cached way out on the “edge” of the network (their caches, and other people’s caches), so its hard to get a video to disappear globally right away. This sometimes angers content owners.

Early on, YouTube leased their hardware.

Spivack went as far as to claim that Twine will be "the first mainstream Semantic Web application" - and it's certainly fair to say that we've heard lots of theory about the Semantic Web ever since Tim Berners-Lee defined it, but as yet there have been very few large scale success stories (if any). Will Twine finally be the Semantic Web app that breaks through? Let's find out more...

First some background: Nova Spivack has an illustrious history in the Semantic Web and AI business, having worked for both AI legend Ray Kurzweil and tech guru Danny Hillis (Thinking Machines). The genesis for Twine, said Spivack, came from an R&D project about 5 years ago, which turned into a research project, then a Series A round with Microsoft co-founder Paul Allen in 2006. As of now the Twine team is 30 people working from San Francisco -- and they're finally ready to unveil their new mainstream Semantic Web product.

What is Twine?

The aim of Twine is to enable people to share knowledge and information. At first glance it is very much like Wikipedia, but there is a whole lot more smarts to the system. Spivack described it to me as "knowledge networking"- i.e. it aims to connect people with each other "for a purpose". It's not based around socializing, but to share and organize information you're interested in. Using Twine, you can add content via wiki functionality (there are many post types), you can email content into the system, and "collect" something (as an object, e.g. a book object). The screenshots below show of this in action -- note that the product itself isn't available just yet, as it's in private testing.

Google currently processes over 20 petabytes of data per day through an average of 100,000 MapReduce jobs spread across its massive computing clusters. The average MapReduce job ran across approximately 400 machines in September 2007, crunching approximately 11,000 machine years in a single month. These are just some of the facts about the search giant's computational processing infrastructure revealed in an ACM paper by Google Fellows Jeffrey Dean and Sanjay Ghemawat.

Twenty petabytes (20,000 terabytes) per day is a tremendous amount of data processing and a key contributor to Google's continued market dominance. Competing search storage and processing systems at Microsoft (Dyrad) and Yahoo! (Hadoop) are still playing catch-up to Google's suite of GFS, MapReduce, and BigTable.

Google processes its data on a standard machine cluster node consisting two 2 GHz Intel Xeon processors with Hyper-Threading enabled, 4 GB of memory, two 160 GB IDE hard drives and a gigabit Ethernet link. This type of machine costs approximately $2400 each through providers such as Penguin Computing or Dell or approximately $900 a month through a managed hosting provider such as Verio (for startup comparisons).

The average MapReduce job runs across a $1 million hardware cluster, not including bandwidth fees, datacenter costs, or staffing.

Summary

The January 2008 MapReduce paper provides new insights into Google's hardware and software crunching processing tens of petabytes of data per day. Google converted its search indexing systems to the MapReduce system in 2003, and currently processes over 20 terabytes of raw web data. It's some fascinating large-scale processing data that makes your head spin and appreciate the years of distributed computing fine-tuning applied to today's large problems.

Since 99% of Google's $6 billion in revenue continues to come from advertising, the Google Enterprise division represents a tiny part of the overall business. This is the group that wants to sell you a little bit of Google in a box—the Search Appliance product line—that embeds a variation of the Google.com search engine software that powers Google.com in a yellow server box or blue blade server that enterprise customers can plug into their own data centers.

The appliance product line includes the entry-level google mini, which starts at $1,995 for a model capable of indexing 100,000 documents. It is often used to provide the search capability for public Web sites, although it is also used internally by small enterprises or departments of larger ones. Beyond that level, the Search Appliance line includes the Google Appliance GB-1001, which can handle up to a million documents; and the GB-5005 and GB-8008, which, when delivered in the form of multiple servers in a rack and working together as a Google File System cluster, can handle many millions of documents. "It really is like a little Google data center in a box," says Matt Glotzbach, head of products for Google Enterprise.

Because the technology is delivered in the form of an appliance, customers aren't supposed to crack open the case and tinker with the technology inside, and they aren't provided with root access to the server operating environment.

Instead, Google provides a set of Web-based administration screens, as well as application programming interfaces for modifying the appliance's behavior. The most significant development on that front is the introduction of the onebox application programming interface, which allows data drawn from other systems that otherwise wouldn't be indexed by the search appliance to be displayed at the top of the search results.

The enterprise version of OneBox is modeled after the feature of Google's public Web site that inserts links to data from weather reports, phone listings or maps into search results when the search engine recognizes a pattern associated with an address, a city name or a phone number in the keywords entered by the user. Similarly, the appliance can be programmed to recognize patterns associated with purchase order numbers or common business queries, and insert links to related data. For example, a search for quarterly sales data could go beyond searching intranet Web content and pop up a link to a more structured data source, such as a Cognos financial analysis application.

Dave Girouard, vice president and general manager of the Google Enterprise business unit, says Google has been beefing up its capabilities to address enterprise requirements, making the appliances easier to install, use and manage.

However, google still needs to do a better job of addressing enterprise requirements, particularly in terms of support, according to Gartner's Whit Andrews, an authority on the evolution of search technology. "It has taken Google a while to recognize that it needs to do business with the enterprise in a different way from how it does business with the advertiser," he says.

Enterprises that have bought the appliances often give positive reports on the value Google delivers for the money and on the ease of setup and administration, Andrews says. But support is another story, according to Andrews, who has talked to customers who say Google's response to problems with the appliances is too often along the lines of, "Yeah, we know about that, we'll get back to you."

Google says it is investing in improved support. Andrews agrees that the support is better, but says it's still not enough.

However, Google can point to many happy customers. Brown Rudnick Berlack Israels, an international law firm based in Boston, got the Google Mini it purchased up and running in about an hour, says Keith Schultz, who manages the firm's Web sites. "We've really had no problems with it at all," Schultz says.

Forging ahead with the same business model for more than 12 years might seem old hat to some in the constantly changing world of information technology, but business customers say Red Hat wears it well.

Brian Goetz (brian.goetz@sun.com), Senior Staff Engineer, Sun Microsystems

19 Jun 2007

The Java™ language contains two intrinsic synchronization mechanisms: synchronized blocks (and methods) and volatile variables. Both are provided for the purpose of rendering code thread-safe. Volatile variables are the weaker (but sometimes simpler or less expensive) of the two -- but also easier to use incorrectly. In this installment of Java theory and practice, Brian Goetz explores some patterns for using volatile variables correctly and offers some warnings about the limits of its applicability.

Volatile variables in the Java language can be thought of as "synchronized lite"; they require less coding to use than synchronized blocks and often have less runtime overhead, but they can only be used to do a subset of the things that synchronized can. This article presents some patterns for using volatile variables effectively -- and some warnings about when not to use them.

Locks offer two primary features: mutual exclusion and visibility. Mutual exclusion means that only one thread at a time may hold a given lock, and this property can be used to implement protocols for coordinating access to shared data such that only one thread at a time will be using the shared data. Visibility is more subtle and has to do with ensuring that changes made to shared data prior to releasing a lock are made visible to another thread that subsequently acquires that lock -- without the visibility guarantees provided by synchronization, threads could see stale or inconsistent values for shared variables, which could cause a host of serious problems.

Volatile variables

Volatile variables share the visibility features of synchronized, but none of the atomicity features. This means that threads will automatically see the most up-to-date value for volatile variables. They can be used to provide thread safety, but only in a very restricted set of cases: those that do not impose constraints between multiple variables or between a variable's current value and its future values. So volatile alone is not strong enough to implement a counter, a mutex, or any class that has invariants that relate multiple variables (such as "start <=end").

You might prefer to use volatile variables instead of locks for one of two principal reasons: simplicity or scalability. Some idioms are easier to code and read when they use volatile variables instead of locks. In addition, volatile variables (unlike locks) cannot cause a thread to block, so they are less likely to cause scalability problems. In situations where reads greatly outnumber writes, volatile variables may also provide a performance advantage over locking.

Conditions for correct use of volatile

You can use volatile variables instead of locks only under a restricted set of circumstances. Both of the following criteria must be met for volatile variables to provide the desired thread-safety:

• Writes to the variable do not depend on its current value.
• The variable does not participate in invariants with other variables.

Basically, these conditions state that the set of valid values that can be written to a volatile variable is independent of any other program state, including the variable's current state.

The first condition disqualifies volatile variables from being used as thread-safe counters. While the increment operation (x++) may look like a single operation, it is really a compound read-modify-write sequence of operations that must execute atomically -- and volatile does not provide the necessary atomicity. Correct operation would require that the value of x stay unchanged for the duration of the operation, which cannot be achieved using volatile variables. (However, if you can arrange that the value is only ever written from a single thread, then you can ignore the first condition.)

Most programming situations will fall afoul of either the first or second condition, making volatile variables a less commonly applicable approach to achieving thread-safety than synchronized. Listing 1 shows a non-thread-safe number range class. It contains an invariant -- that the lower bound is always less than or equal to the upper bound.

                
@NotThreadSafe 
public class NumberRange {
    private int lower, upper;

    public int getLower() { return lower; }
    public int getUpper() { return upper; }

    public void setLower(int value) { 
        if (value > upper) 
            throw new IllegalArgumentException(...);
        lower = value;
    }

    public void setUpper(int value) { 
        if (value < lower) 
            throw new IllegalArgumentException(...);
        upper = value;
    }
}

Because the state variables of the range are constrained in this manner, making the lower and upper fields volatile would not be sufficient to make the class thread-safe; synchronization would still be needed. Otherwise, with some unlucky timing, two threads executing setLower and setUpper with inconsistent values could leave the range in an inconsistent state. For example, if the initial state is (0, 5), and thread A calls setLower(4) at the same time that thread B calls setUpper(3), and the operations are interleaved just wrong, both could pass the checks that are supposed to protect the invariant and end up with the range holding (4, 3) -- an invalid value. We need to make the setLower() and setUpper() operations atomic with respect to other operations on the range -- and making the fields volatile can't do this for us.

Performance considerations

The primary motivation for using volatile variables is simplicity: In some situations, using a volatile variable is just simpler than using the corresponding locking. A secondary motivation for using volatile variables is performance: In some situations, volatile variables may be a better-performing synchronization mechanism than locking.

It is exceedingly difficult to make accurate, general statements of the form "X is always faster than Y," especially when it comes to intrinsic JVM operations. (For example, the VM may be able to remove locking entirely in some situations, which makes it hard to talk about the relative cost of volatile vs. synchronized in the abstract.) That said, on most current processor architectures, volatile reads are cheap -- nearly as cheap as nonvolatile reads. Volatile writes are considerably more expensive than nonvolatile writes because of the memory fencing required to guarantee visibility but still generally cheaper than lock acquisition.

Unlike locking, volatile operations will never block, so volatiles offer some scalability advantages over locking in the cases where they can be used safely. In cases where reads greatly outnumber writes, volatile variables can often reduce the performance cost of synchronization compared to locking.

Patterns for using volatile correctly

Many concurrency experts tend to guide users away from using volatile variables at all, because they are harder to use correctly than locks. However, some well-defined patterns exist, which, if you follow them carefully, can be used safely in a wide variety of situations. Always keep in mind the rules about the limits of where volatile can be used -- only use volatile for state that is truly independent of everything else in your program -- and this should keep you from trying to extend these patterns into dangerous territory.

Pattern #1: status flags

Perhaps the canonical use of volatile variables is simple boolean status flags, indicating that an important one-time life-cycle event has happened, such as initialization has completed or shutdown has been requested.

Many applications include a control construct of the form, "While we're not ready to shut down, do more work," as shown in Listing 2:

                
volatile boolean shutdownRequested;

...

public void shutdown() { shutdownRequested = true; }

public void doWork() { 
    while (!shutdownRequested) { 
        // do stuff
    }
}

It is likely that the shutdown() method is going to be called from somewhere outside the loop -- in another thread -- and as such, some form of synchronization is required to ensure the proper visibility of the shutdownRequested variable. (It might be called from a JMX listener, an action listener in the GUI event thread, through RMI, through a Web service, and so on.) However, coding the loop with synchronized blocks would be much more cumbersome than coding it with a volatile status flag as in Listing 2. Because volatile simplifies the coding, and the status flag does not depend on any other state in the program, this is a good use for volatile.

One common characteristic of status flags of this type is that there is typically only one state transition; the shutdownRequested flag goes from false to true and then the program shuts down. This pattern can be extended to state flags that can change back and forth, but only if it is acceptable for a transition cycle (from false to true to false) to go undetected. Otherwise, some sort of atomic state transition mechanism is needed, such as atomic variables.

Pattern #2: one-time safe publication

The visibility failures that are possible in the absence of synchronization can get even trickier to reason about when writing to object references instead of primitive values. In the absence of synchronization, it is possible to see an up-to-date value for an object reference that was written by another thread and still see stale values for that object's state. (This hazard is the root of the problem with the infamous double-checked-locking idiom, where an object reference is read without synchronization, and the risk is that you could see an up-to-date reference but still observe a partially constructed object through that reference.)

One technique for safely publishing an object is to make the object reference volatile. Listing 3 shows an example where during startup, a background thread loads some data from a database. Other code, when it might be able to make use of this data, checks to see if it has been published before trying to use it.

                
public class BackgroundFloobleLoader {
    public volatile Flooble theFlooble;

    public void initInBackground() {
        // do lots of stuff
        theFlooble = new Flooble();  // this is the only write to theFlooble
    }
}

public class SomeOtherClass {
    public void doWork() {
        while (true) { 
            // do some stuff...
            // use the Flooble, but only if it is ready
            if (floobleLoader.theFlooble != null) 
                doSomething(floobleLoader.theFlooble);
        }
    }
}

Without the theFlooble reference being volatile, the code in doWork() would be at risk for seeing a partially constructed Flooble as it dereferences the theFlooble reference.

A key requirement for this pattern is that the object being published must either be thread-safe or effectively immutable (effectively immutable means that its state is never modified after its publication). The volatile reference may guarantee the visibility of the object in its as-published form, but if the state of the object is going to change after publication, then additional synchronization is required.

Pattern #3: independent observations

Another simple pattern for safely using volatile is when observations are periodically "published" for consumption within the program. For example, say there is an environmental sensor that senses the current temperature. A background thread might read this sensor every few seconds and update a volatile variable containing the current temperature. Then, other threads can read this variable knowing that they will always see the most up-to-date value.

Another application for this pattern is gathering statistics about the program. Listing 4 shows how an authentication mechanism might remember the name of the last user to have logged on. The lastUser reference will be repeatedly used to publish a value for consumption by the rest of the program.

                
public class UserManager {
    public volatile String lastUser;

    public boolean authenticate(String user, String password) {
        boolean valid = passwordIsValid(user, password);
        if (valid) {
            User u = new User();
            activeUsers.add(u);
            lastUser = user;
        }
        return valid;
    }
} 

This pattern is an extension of the previous one; a value is being published for use elsewhere within the program, but instead of publication being a one-time event, it is a series of independent events. This pattern requires that the value being published be effectively immutable -- that its state not change after publication. Code consuming the value should be aware that it might change at any time.

Pattern #4: the "volatile bean" pattern

The volatile bean pattern is applicable in frameworks that use JavaBeans as "glorified structs." In the volatile bean pattern, a JavaBean is used as a container for a group of independent properties with getters and/or setters. The rationale for the volatile bean pattern is that many frameworks provide containers for mutable data holders (for instance, HttpSession), but the objects placed in those containers must be thread safe.

In the volatile bean pattern, all the data members of the JavaBean are volatile, and the getters and setters must be trivial -- they must contain no logic other than getting or setting the appropriate property. Further, for data members that are object references, the referred-to objects must be effectively immutable. (This prohibits having array-valued properties, as when an array reference is declared volatile, only the reference, not the elements themselves, have volatile semantics.) As with any volatile variable, there may be no invariants or constraints involving the properties of the JavaBean. An example of a JavaBean obeying the volatile bean pattern is shown in Listing 5:

                
@ThreadSafe
public class Person {
    private volatile String firstName;
    private volatile String lastName;
    private volatile int age;

    public String getFirstName() { return firstName; }
    public String getLastName() { return lastName; }
    public int getAge() { return age; }

    public void setFirstName(String firstName) { 
        this.firstName = firstName;
    }

    public void setLastName(String lastName) { 
        this.lastName = lastName;
    }

    public void setAge(int age) { 
        this.age = age;
    }
}

Advanced patterns for volatile

The patterns in the previous section cover most of the basic cases where the use of volatile is sensible and straightforward. This section looks at a more advanced pattern where volatile might offer a performance or scalability benefit.

The more advanced patterns for using volatile can be extremely fragile. It is critical that your assumptions be carefully documented and these patterns strongly encapsulated because very small changes can break your code! Also, given that the primary motivation for the more advanced volatile use cases is performance, be sure that you actually have a demonstrated need for the purported performance gain before you start applying them. These patterns are trade-offs that give up readability or maintainability in exchange for a possible performance boost -- if you don't need the performance boost (or can't prove you need it through a rigorous measurement program), then it is probably a bad trade because you're giving up something of value and getting something of lesser value in return.

Pattern #5: The cheap read-write lock trick

By now, it should be well-known that volatile is not strong enough to implement a counter. Because ++x is really shorthand for three operations (read, add, store), with some unlucky timing it is possible for updates to be lost if multiple threads tried to increment a volatile counter at once.

However, if reads greatly outnumber modifications, you can combine intrinsic locking and volatile variables to reduce the cost on the common code path. Listing 6 shows a thread-safe counter that uses synchronized to ensure that the increment operation is atomic and uses volatile to guarantee the visibility of the current result. If updates are infrequent, this approach may perform better as the overhead on the read path is only a volatile read, which is generally cheaper than an uncontended lock acquisition.

The reason this technique is called the "cheap read-write lock" is that you are using different synchronization mechanisms for reads and writes. Because the writes in this case violate the first condition for using volatile, you cannot use volatile to safely implement the counter -- you must use locking. However, you can use volatile to ensure the visibility of the current value when reading, so you use locking for all mutative operations and volatile for read-only operations. Where locks only allow one thread to access a value at once, volatile reads allow more than one, so when you use volatile to guard the read code path, you get a higher degree of sharing than you would were you to use locking for all code paths -- just like a read-write lock. However, bear in mind the fragility of this pattern: With two competing synchronization mechanisms, this can get very tricky if you branch out beyond the most basic application of this pattern.

Summary

Volatile variables are a simpler -- but weaker -- form of synchronization than locking, which in some cases offers better performance or scalability than intrinsic locking. If you follow the conditions for using volatile safely -- that the variable is truly independent of both other variables and its own prior values -- you can sometimes simplify code by using volatile instead of synchronized. However, code using volatile is often more fragile than code using locking. The patterns offered here cover the most common cases where volatile is a sensible alternative to synchronized. Following these patterns -- taking care not to push them beyond their limits -- should help you safely cover the majority of cases where volatile variables are a win.