Prev: paging-smaller-tables Next: beyond-physical-memory-policies
We previously assumed that address space fits in main memory. Now, lets make it work on disk.
The addition of swap space allows the OS to support the illusion that there’s a large virtual memory for concurrently running processes.
We reserve some space on disk to store pages that hold processes. The OS needs to remember the disk address of a given page.
Since we can have pages on disk as well as in memory, we’ll also add a bit to a page that states whether or not the page is present in memory – if not, the OS invokes a page-fault handler, which goes to the OS to find the given page on swap space on disk.
Page-Fault handlers are normally run in software – this is because hardware doesn’t necessarily know where in the Page table entry to find the address, and I/O requests are slow anyway, so there’s not much performance loss – as well, the OS is well equipped to handle those requests and interleave them efficiently.
If memory is full, we need to find a policy to kick a page out to main memory – but of course, that has a performance penalty – if the wrong page is sent to main memory, a program could run close to 100,000 times slower. Therefore, its important to have a good policy, which is discussed in the next chapter.
The control flow looks like this: If the page is present in valid, the TLB can find the page, and retry the instruction.
If the page is not present in physical memory, then the OS goes to main memory, after being notified by a page fault handler.
Otherwise, the OS waits for main memory to fetch the page and then gets a TLB cache hit to service the request.
The OS does bookkeeping in the background – it has two markers, a high watermark and a low watermark – the low watermark is when the OS decides to free memory until it hits the high watermark. This is done by the swap daemon or page daemon.
This allows for improved performance, since it allows for grouping of work and batching during a better time for the hardware, and allows software to act like its working faster.
Prev: paging-smaller-tables Next: beyond-physical-memory-policies