When do we write a page from physical memory back to the disk?

In general, caches have two broad types of writing policies. One approach is a write-through cache. In this case, when a value in the cache is written, it is immediately written to the backing store as well (in this case, the disk). The cache and backing store are always synchronized in this case, but this can be very slow. The other main approach is a write-back cache. In this case, the backing store and the cache are sometimes out of sync, but this approach is much faster. This is what is used with paging, for obvious speed reasons.

When a page is loaded from the disk to physical memory, it is initially clean, i.e. the copy in physical memory matches the copy on disk. If the copy in memory is ever changed, then its page-table entry is marked dirty, and it will need to be written back to the disk later. When physical memory ?lls up, and a non-resident page is requested, then the OS needs to select a page to evict, to make room for the new page. The evicted page is called the victim, and is saved to the so-called "swap" space.

The swap space is a separate region of the disk from the ?le system, and the size of the swap space limits the total virtual address space of all programs put together (though in practice, there is a lot of memory shared between processes, for instance shared libraries). There are a variety different strategies for choosing which page to evict, with tradeoffs for each strategy. These strategies will be discussed later. One thing to note is that evicting a clean page is fast, since it doesn't need to be written back to the disk. A second note is that to speed up the process of evicting pages, the OS can write dirty pages back to disk as a background task. In this way, more pages will be clean and can therefore be evicted a lot more quickly, when it is time to do so.