Virtual Memory Management in the VAX/VMS OS

One-line summary: This paper presents a virtual memory system intended to run with low overhead and I/O on a wide variety of hardware configurations.

Overview/Main Points

• VM Setup/Details:
  • 32-bit virtual address space : 21-bit page number, 9 bit offset (512 byte pages).
  • Top two bits divide the system into system and process space.
  • Page tables are contiguous arrays of 32-bit page-table entries, including valid bit, protection field, modify bit, physical frame mapping, and os-specific bits.
  • The page table for the system space is accesses through a base register, and is never affected by contect switches. The tables for the process space are virtually addressed in the system space. Translation of virtual addresses take two accesses (one to access the system page table and get the physical process page table, and one to get the physical address from the process page table).
  • Translation buffer for caching tranlations (no size info...)
  • Address space has pointers to system service routines (to allow for later modifications to the routines). Memory system does not have a separate context, so all of these routines must be re-entrant.
  • First page in process space is unaccessible, to prevent uninitialized pointer references.

• Pager:
  • Each process has a limited resident set of pages. When this set is full, the process must release a page to fault in a new page.
  • FIFO is used to select pages for replacement, based on their assumption that low processor usage was more important than increased memory usage. (How has this changed to the present?)
  • Free and modified lists provide page caching. Removed pages put on these lists can be recovered cheaply when re-touched in the time period before these pages are reused.
  • Reads and writes to swap space are clustered. When reading, contiguous pages on disk from the same virtual space are loaded (What is the overhead on this? No statistics...). When writing, virtually contiguous pages are written in contiguous batches on disk. Held pages are often either paged back or invalidated (through program termination).
  • Process page tables are part of their resident set, but are only paged out if they have no active page entries (What prevents a large program from filling its pages with page table pages and thrashing?).

• Swapper:

  Processes are swapped as units (after waiting for any operations in progress to complete).

Relevance

• Caching of pages to reduce disk I/Os for pages recently removed from a process's resident set.
• Clustering (simulates the efficientcy of larger page transfers).
• Process-local replacement, to reduce the effect of process' paging activities on each other.

Flaws

• Lack of statistics/numbers. More info on how the system performed under different loads and resident set sizes would have been useful.