Allocation of Page Frames

1
Allocation of Page Frames

• How many page frames should be allocated to a
  process?
• Should allocate according to size of the working
  set of the process
• Remember that working set is set of pages that
  process has referenced in the last N seconds or
  last N references (not easily known at startup)
• Allocating a process fewer pages than its working
  set can quickly lead to page faults
• One way to prevent many page faults is to not
  schedule a process unless it can be allocated
  enough page frames for its working set

2
Allocation of Page Frames

• Local vs. global allocation
• Local when a process needs more pages, evict
  its own pages (not another process)
• Can lead to wasted memory if a process working
  set decreases in size
• When a process thrashes, it will not cause other
  processes to thrash
• Global on a page fault, evict a page from any
  process page set
• Better for dynamic working set which may grow or
  shrink
• Domino-style thrashing may occur one process
  page faults, evicting another process pages, and
  when this process runs, it will evict yet another
  process pages, etc.

3
Allocation Policies

• Periodically determine number of running
  processes and allocate each an equal share
  (hybrid of local global)
• Allocate number of pages proportional to size of
  process
• Waste and thrashing may still occur

4
Allocation of Page Frames

• Page fault frequency measure of page fault rate
  of a process
• Too low take pages from that process
• Too high allocate additional page frames to the
  process
• Too many processes have high page fault frequency
  swap out one or more of these processes (which
  process to swap out determined by scheduling
  algorithm)
• May also use process priority to affect page
  allocation (thereby, affecting page replacement)

5
Relationship of Allocation to Replacement Policy

• FIFO and LRU can be run either globally or
  locally on the process pages
• Working set makes sense only as a local policy or
  to initialize the number of pages given to a
  process

6
Policies to Prevent Thrashing

• Adopt a local allocation policy
• Do not schedule a process unless its working set
  of pages is in memory (implies prepaging)
• Use Page Fault Frequency approach

7
Page Size

• Want to set page size to reduce internal
  fragmentation
• Smaller page size leads to more pages leads to
  larger page table
• What is optimal page size?
• p page size ?
• s average process size,
• e size of each page table entry
• s/p average number of pages per process
• (s/p)e space taken up by average process in
  page table
• p/2 average wasted memory in last page of
  process due to internal fragmentation
• overhead (s/p)e p/2
• Find optimal page size by taking derivative and
  setting to 0
• p 2se

8
Page Fault Handling

• If present/absent bit 0, means page is not
  addressable by the process (protected addresses)
  or it is not in memory
• Consult backing store map to distinguish between
  the cases
• Absent from map means illegal address
• Otherwise, backing store map indicates where to
  find the page on disk

9
Backing Store

• Swap area on disk where page that is evicted is
  placed
• How should space in swap area be allocated to a
  process?
• When process started, reserve chunk in swap space
  equal to process size
• Either initialize chunk with copy of entire
  process, or load entire process into memory and
  let it be paged out
• Disadvantage size of process may grow. Better
  to have separate swap areas for text, data, and
  stack
• Do not reserve space in swap area in advance
• Allocate disk space for each page when it is
  swapped out (de-allocate when swapped in)
• Disadvantage disk address needed for each page
  (instead of per process)

10
Backing Store Map 2 Implementations

• Store disk address in PTE
• Problems
• Can be wasteful since page table is stored in
  memory
• PTE structure is influenced by hardware, so need
  hardware cooperation to store disk addresses
• Disk addresses are big (device block number)
  and OS-specific
• Would need to have disk address for every page
  not in memory, so more difficult to compress page
  tables
• Use separate data structure
• Associates addressable regions (text, stack,
  data) of virtual address space with starting
  block number on disk
• Each region of virtual address space is stored
  contiguously on disk
• Backing store map data structure stored in memory
  has only one entry per region

11
OS Involvement with Paging

• Process creation
• OS determines program and data size
• Allocates and initializes page table
• Allocates swap area on disk to store page table
  when process is swapped out
• Record page table and swap area in process table
• Possibly preload pages
• Process scheduled for execution
• Reset MMU for process
• Flush TLB
• Copy page table or start/end addresses to
  hardware registers
• Process termination
• Release page table, page frames, and disk swap
  space

12
OS Involvement with Paging

• Page fault
• Determine which virtual address caused page fault
• Compute which table is needed to locate it on
  disk
• Possibly evict a page to make room in main memory
• Back-up program counter to point to faulting
  instruction so that it can be executed again (but
  this time, the needed page is in memory)

13
Unix Paging Policy

• Demand paging
• Page replacement algorithm
• Maintain a certain number of free page frames
  (within a min/max range)
• Swaps out pages when number of free pages is
  below min
• Some Unix variants uses 2-handed clock

14
Linux Paging Policy

• Demand paging
• Maintain a certain range of free page frames
• Each process on a 32-bit machine is given 3 GB of
  virtual address space and 1 GB reserved for page
  tables and other kernel data
• 3-level page table
• Kernel is never paged out
• Buddy system for memory partitioning

15
Windows Paging Policy

• Demand paging without pre-paging
• Maintain a certain number of free page frames
• For 32-bit machine, each process has 4 GB of
  virtual address space
• Backing store disk space is not assigned to
  page until it is paged out
• Uses working sets (per process)
• Consists of pages mapped into memory and can be
  accessed without page fault
• Has min/max size range that changes over time
• If page fault occurs and working set lt min, add
  page
• If page fault occurs and working set gt max, evict
  page from working set and add new page
• If too many page faults, then increase size of
  working set
• When evicting pages,
• Evict from large processes that have been idle
  for a long time before small active processes
• Consider foreground process last