Cache coherence

[jasperzhong]

https://en.wikipedia.org/wiki/Cache_coherence

write-through vs write-back

[jasperzhong]

当系统写数据到cache的时候，之后肯定要把这个数据写入到backing store.

write policy有两个基本方式:

• write-through（透写）: 同时写入cache和backing store.
• write-back（回写）： 开始只写入cache. 直到修改的内容要被另一个cache block替换掉的时候才写到backing store. (lazy write)

所以write-back策略下，一次read miss会导致两次memory accesses: 一个是把要替换的data写入到backing store, 一个是retrieve需要的data.

• write-back一般用write allocate (也叫fetch on write): 就是write miss的时候把数据load到cache, 然后write-hit. 和read miss差不多.
• write-through一般用no-write allocate策略 (也叫write-no-allocate或者write around): 就是write miss的时候不会load数据，而是直接写入到backing store. 只有read miss的时候才会load数据到cache.

[jasperzhong]

wikipedia上的这两个图总结了上面流程:

A write-through cache with no-write allocation:

A write-back cache with write allocation:

顺带提一句，这些都是用硬件完成的.

[jasperzhong]

脏数据（dirty data）指的是缓存上还没有写入到backing store的数据.