Skip to content

Latest commit

 

History

History
59 lines (48 loc) · 2.42 KB

Memory Resource Management in VMware ESX Server.md

File metadata and controls

59 lines (48 loc) · 2.42 KB
Raw

#Memory resource management in VMware ESX server

Key ideas

  • Ballooning to allow memory overcommittment over all VMs
  • Idle memory taxing
  • Content-based page sharing
  • I/O page remapping

Memory virtualization

  • VM has the illusion that "physical addresses" are zero based
  • pmap data structure maps "physical page" number with machine page number
    • VM instructions that would modify memory are intercepted and remapped
  • Shadow page tables map virtual pages to machine pages to avoid 1 level of indirection ("physical pages")

Page replacement

  • When memory is overcomitted, host has to reclaim space
  • Problem: OS in VM may reclaim pages that were just paged out

Ballooning

  • Balloon module in guest OS allocates "physical pages"
  • When inflated - OS realizes how much memory it has available (balloon size)
  • When deflated - OS allocates balloon pages to more sensible uses
  • If balloon driver is unavailable, ESX will swap in disk

Sharing memory

  • If VMs run the same OS or same application, share the actual machine pages
  • pmap of each VM makes the mapping to the same machine page

Transparent page shareing

  • Once copies are identified, map "physical pages" to machine pages
  • Disco's approach, ESX improves it with...

Content-based page sharing

  • Hashing to match memory pages between all domains
  • If page contents are found in hash table, success
    • Make machine page shared
  • If page contents are not found in hash table
    • Mark it as a hint entry, in case of future matches

Implementation

  • Global hash table for all scanned pages
  • Shared frame: hash value, MPN, references count
  • Hint frame: hash value, MPN, PPN, VM identifier
  • Pages are scanned randomly to check equality

Share-based allocation

  • Resource rights are determined by shares, owned by clients
  • Memory is revoked based on whether client owns enough shares per page

Idle memory tax

  • Clients are charged more for idle pages than used ones
  • Ensures memory that is idle is reclaimed first
  • Host does statistical sampling of pages to measure idleness
  • A small buffer of 25% of idle pages is left in VM to allow for rapid mem increases

I/O page remapping

  • ESX keeps track of "hot" pages that are frequently used for I/O in the guests
  • Since I/O happens on low memory normally, ESX maps those "physical pages" to low memory
  • * low memory is memory permanently mapped into kernel address space