Why superpages?

• Used to reduce indirection in table walking…
  • …which can help decrease frequency of page table access
• Primarily aimed at reducing TLB pressure
  • Making memory access more efficient by utliising larger page sizes

Translation Lookaside Buffer

• Acts as a cache for the page table.
  • Stores recently access virtual-to-physical address translations
• When a memory access is requestion, TLB is checked first to see if translation is already available
  • Helps speed up memory access process
• If required access not found in the TLB, a a TLB miss

TLB Miss

{width=250px}

• Happens when TLB does not contain the required translation for virtual-to-physical address mapping
• When a TLB miss happens, hardware page-table walker or OS must retrieve the translation from the page table
• If no entry in the TLB or the page table, Page fault

Page Fault

{width=450px}

• Page fault handler triggered to handle the above situation by…
  • Populating memory
  • Updating virtual memory mappings
  • Potentially fetching data from secondary storage (disk, network, etc.)
• This process ensures the required data is brought into memory and mapping is establish
  • Done to allow program to continue execution
• Goal: Keep page faults to a minimum
  • Page replacement
  • Page fetching
  • Buffering
  • Resident size management
  • etc.

Page Fetch Strategies

• Demand paging: Pages fetches only when needed
  • Initially leads to many page faults, but decreases overtime due to temporal locality
• Prepaging: Preloads memory by exploiting secondary disk structure
  • One page fault on startup
• Which one is better?:
  • Prepaging is generally considered better, as it preloads memory and incurs only one page fault on startup, reducing initial delay in accessing pages.

Replacement Policies

• When loading from secondary storage, what should be replaced?
  • Replacement policies dictate which page or frame should be replaced when a new page needs to be brought into memory
• Consider frame locking and R/W, U/S bits.

Optimal Strategy (Belady’s Algorithm):

• Evicts the page that will be used in the future
• Not always achievable due to lack of future knowledge of memory access patterns.

Least Recently Used (LRU):

• Replaces the least recently accessed page when a new page needs to be brought into memory
• Good for capturing temporal locality
• Almost as good as Belady in minimising page faults
• However, requires significant overhead to track usage of pages over time at each memory access

First-In-First-Out (FIFO)

• Replaces the page that’s been in memory the longest when a new page needs to be loaded
• Not a great strategy
  • Some frames loaded early are used often
  • $\therefore$ can result in many page faults
• Does not consider the frequency of page accesses

Clock Strategy:

• Treats available pages as circular buffer
• Associates a use bit with each page
  • When a page is accessed, set “A” bit
  • When replacement is needed, scan buffer
    • If “A” bit set, clear bit and move on (done by OS)
    • Otherwise, page’s frame is replaced
• Better Why?:
  • Protects frequently-used pages efficiently
  • Reduces overhead in replacing unmodified pages
• Can also track dirty bit

Page Buffering:

• An add-on technique to replacement algorithm
• OS maintains two pools of pages
  • Free, not modified
  • Free, modified
• When a plage is “replaced”, it is placed in one of these pools first
• Checks pool first when page requested

Resident Set Management:

• Managing portion of process’ memory physically held in main memory
• VM system decides size of each process:
  • Small: Lots of processes, higher page faults
  • Large: Few processes, low multiprogramming
• Can be done through different modes of allocation:
  • Fixed allocation:
    • Number of frames fixed at process load time
    • Can be equal or proportional allocation
    • Page faults bump a page from the same process
  • Variable allocation:
    • Frames per process can vary over time
      • Increase if page fault rate is high
      • Decrease if slow
    • Tracking page fault rates increases overhead by a little

Cleaning

• Periodically writing out modified pages to ensure the most up-to-date data is stored on disk
• Adds overhead of a page fault, but helps maintain data integrity and consistency
• Essential to ensure modified pages are saved before replacement, preventing data loss or corruption
• Demand Cleaning:
  • Write out modifications as pages are replaced
  • Adds to the overhead of a page fault
• Pre-cleaning:
  • Periodically write out modified pages
  • Allows us to write out in batches
• Both can be used with page buffering