Beyond Physical Memory: Mechanisms

• Until now, we have assumed that memory used by every process fits in physical memory (all at once). That’s not realistic.
• Key idea: If no frame is available, move an unused frame to physical memory.
  • This is more or less the same idea as cache: Keep what you are using “close” and put other stuff “far away”.
• Differences?
  • Cache is small relative to main memory (8KB for L1 compared to 10s of GB).
  • Today, we only swap out a small percentage of RAM (e.g., 3GB out of 16GB).
  • RAM is about 100x slower than cache.
  • Disk is about 10,000 times slower than RAM.
• Compare “single address space” system to old-school Overlays that the programmer had to manage by hand.
• Swap space: Space on disk for vm pages.
• Idea: Instead of swapping code pages out to the specific swap space, just dump them and reload them from disk if necessary.
  • Caveat?
  • You can’t delete a program while it is running.
• Add a present bit to the page table.
  • Set the present bit to 0 if the page is not in a page frame, but on disk instead.
• A page fault happens when a page is not present.
• A page fault is handled by the operating system, not hardware
  • There are a lot of steps involved in doing the disk I/O and loading the resulting data into RAM.
• Often the page table entry can contain the location on disk if page isn’t present.
• Process is blocked while page fault is handled. (Let’s other processes get work done.)
• What are the three possibilities with TLB miss:
  1. Present and valid (Just grab frame number from the page table)
  2. Valid and not present (Need to fetch page from swap space)
  3. Invalid

Optimization: Keep some page frames free

• Waiting until memory is full makes page faults really slow: We’d need to wait for page to be written out, then new page read in.
• Idea: Preemptively write pages out.
• When number of free pages goes below the low water mark, start writing pages out to disk.
  • (This can even be done when the disk isn’t otherwise busy, giving priority to I/O from running processes).
• Stop writing pages out when a high water mark is reached.
• To improve efficiency, write out a group of related pages
  • Likely to be evicted around the same time
  • Larger disk writes are more efficient.
• Key idea: Just because a page is written out, does not mean it is no longer present.
  • If the page is referenced before the page frame is needed, it is “saved”. No page fault.
  • This is the idea of “major” and “minor” page faults from recent lab.