综上所述，这个问题的正确答案是: 如果代码在2.0运行时或更高版本中运行，volatile关键字几乎不需要，如果不必要地使用，弊大于利。也就是说，永远不要用它。但是在运行时的早期版本中，需要对静态字段进行适当的双重检查锁定。特别是具有静态类初始化代码的静态字段。

CLR喜欢优化指令，所以当你在代码中访问一个字段时，它可能并不总是访问该字段的当前值(它可能来自堆栈等)。将字段标记为volatile可以确保该指令可以访问该字段的当前值。当程序中的并发线程或操作系统中运行的其他代码可以修改该值(在非锁定场景中)时，这很有用。

您显然失去了一些优化，但它确实使代码更加简单。

如果你想稍微了解一下volatile关键字的功能，可以考虑以下程序(我使用的是DevStudio 2005):

#include <iostream>
void main()
{
  int j = 0;
  for (int i = 0 ; i < 100 ; ++i)
  {
    j += i;
  }
  for (volatile int i = 0 ; i < 100 ; ++i)
  {
    j += i;
  }
  std::cout << j;
}

使用标准的优化(发布)编译器设置，编译器创建以下汇编器(IA32):

void main()
{
00401000  push        ecx  
  int j = 0;
00401001  xor         ecx,ecx 
  for (int i = 0 ; i < 100 ; ++i)
00401003  xor         eax,eax 
00401005  mov         edx,1 
0040100A  lea         ebx,[ebx] 
  {
    j += i;
00401010  add         ecx,eax 
00401012  add         eax,edx 
00401014  cmp         eax,64h 
00401017  jl          main+10h (401010h) 
  }
  for (volatile int i = 0 ; i < 100 ; ++i)
00401019  mov         dword ptr [esp],0 
00401020  mov         eax,dword ptr [esp] 
00401023  cmp         eax,64h 
00401026  jge         main+3Eh (40103Eh) 
00401028  jmp         main+30h (401030h) 
0040102A  lea         ebx,[ebx] 
  {
    j += i;
00401030  add         ecx,dword ptr [esp] 
00401033  add         dword ptr [esp],edx 
00401036  mov         eax,dword ptr [esp] 
00401039  cmp         eax,64h 
0040103C  jl          main+30h (401030h) 
  }
  std::cout << j;
0040103E  push        ecx  
0040103F  mov         ecx,dword ptr [__imp_std::cout (40203Ch)] 
00401045  call        dword ptr [__imp_std::basic_ostream<char,std::char_traits<char> >::operator<< (402038h)] 
}
0040104B  xor         eax,eax 
0040104D  pop         ecx  
0040104E  ret              

Looking at the output, the compiler has decided to use the ecx register to store the value of the j variable. For the non-volatile loop (the first) the compiler has assigned i to the eax register. Fairly straightforward. There are a couple of interesting bits though - the lea ebx,[ebx] instruction is effectively a multibyte nop instruction so that the loop jumps to a 16 byte aligned memory address. The other is the use of edx to increment the loop counter instead of using an inc eax instruction. The add reg,reg instruction has lower latency on a few IA32 cores compared to the inc reg instruction, but never has higher latency.

Now for the loop with the volatile loop counter. The counter is stored at [esp] and the volatile keyword tells the compiler the value should always be read from/written to memory and never assigned to a register. The compiler even goes so far as to not do a load/increment/store as three distinct steps (load eax, inc eax, save eax) when updating the counter value, instead the memory is directly modified in a single instruction (an add mem,reg). The way the code has been created ensures the value of the loop counter is always up-to-date within the context of a single CPU core. No operation on the data can result in corruption or data loss (hence not using the load/inc/store since the value can change during the inc thus being lost on the store). Since interrupts can only be serviced once the current instruction has completed, the data can never be corrupted, even with unaligned memory.

Once you introduce a second CPU to the system, the volatile keyword won't guard against the data being updated by another CPU at the same time. In the above example, you would need the data to be unaligned to get a potential corruption. The volatile keyword won't prevent potential corruption if the data cannot be handled atomically, for example, if the loop counter was of type long long (64 bits) then it would require two 32 bit operations to update the value, in the middle of which an interrupt can occur and change the data.

因此，volatile关键字只适用于小于或等于本机寄存器大小的对齐数据，这样操作总是原子的。

volatile关键字被设想用于IO操作，其中IO将不断变化，但有一个恒定的地址，例如内存映射的UART设备，编译器不应该一直重用从地址中读取的第一个值。

如果要处理大数据或有多个cpu，则需要更高级别(OS)的锁定系统来正确处理数据访问。

多个线程可以访问一个变量。 最新的更新将在变量上

我认为没有比Eric Lippert更好的人来回答这个问题了(在原文中强调):

In C#, "volatile" means not only "make sure that the compiler and the jitter do not perform any code reordering or register caching optimizations on this variable". It also means "tell the processors to do whatever it is they need to do to ensure that I am reading the latest value, even if that means halting other processors and making them synchronize main memory with their caches". Actually, that last bit is a lie. The true semantics of volatile reads and writes are considerably more complex than I've outlined here; in fact they do not actually guarantee that every processor stops what it is doing and updates caches to/from main memory. Rather, they provide weaker guarantees about how memory accesses before and after reads and writes may be observed to be ordered with respect to each other. Certain operations such as creating a new thread, entering a lock, or using one of the Interlocked family of methods introduce stronger guarantees about observation of ordering. If you want more details, read sections 3.10 and 10.5.3 of the C# 4.0 specification. Frankly, I discourage you from ever making a volatile field. Volatile fields are a sign that you are doing something downright crazy: you're attempting to read and write the same value on two different threads without putting a lock in place. Locks guarantee that memory read or modified inside the lock is observed to be consistent, locks guarantee that only one thread accesses a given chunk of memory at a time, and so on. The number of situations in which a lock is too slow is very small, and the probability that you are going to get the code wrong because you don't understand the exact memory model is very large. I don't attempt to write any low-lock code except for the most trivial usages of Interlocked operations. I leave the usage of "volatile" to real experts.

欲进一步阅读，请参阅:

理解低锁技术在多线程应用中的影响 再见不稳定

从MSDN: volatile修饰符通常用于由多个线程访问而不使用lock语句序列化访问的字段。使用volatile修饰符可确保一个线程检索到另一个线程写入的最新值。