VMworld 2014/Extreme Performance Series: Understanding Virtualized Memory Management Performance

Extreme Performance Series: Understanding Virtualized Memory Management Performance (INF1420)

Motivations

configured memory size:

• too small impacts performance
• too large impacts overhead

number of VMs per host:

• too many impacts performance
• too few wastes host memory

memory reclamation method:

• properly implemented, will minimally impact performance

Layers

Memory Layers

• App (ex. Java)
• VM (Guest OS)
• Host (ESX)

Each layer does memory management differently

Terminology

• Memory Size
  • Free Memory
  • Allocated Memory
    • Active Memory
    • Idle Memory

Memory Undercommit

• Sum of all VM memory is less than host memory
• No reclamation of guest memory, even if VM has free/idle memory

Memory Overcommit

• Sum of all VM memory is more than host memory
• ESX may map only a subset of VM memory

Active, Idle, Cache, Free, Ballooned

entitlement

Memory Entitlement and Reclamation

• compute memory entitlement for each vm, then reclaim guest memory if entitlement is less than consumed
• computed on Reservation, Limit, Shares, and Memory demand
• esx can only classify memory broadly as active and idle
• sampling baed demand estimation is reasonably accurate

Parameters:

• configured memory size
• memory reservation
• memory limit
• memory shares
• idle memory ("idle memory tax")

Memory Reclamation Techniques

Memory Reclamation Techniques

• Transparent page sharing - de-dup identical memory pages (low cpu cost technique, uses content hasing)
• Ballooning
• Compression
• Swapping

Performance Implications

Designed to handle burst memory pressure well

Constant memory pressure should be handled by DRS/vMotion etc.

Monitoring Tools

• vCenter performance chart
• esxtop
• memstats
• vCenter Operations (vCOps) - add on software
  • custom UI: http://vcops-host/vcops-custom

VM Sizing

• guard against "active memory" reclamation
• sized larger than highest demand during peak loads

Page Sharing & Large Page

• Savings can be significant for homogeneous VMs (like VDI)
• "Preventers":
  • ASLR (Address space Layout Randomizaiton) - for security
  • Super fetching (proactive caching)
  • Host large pages (ESXi does not share large pages)
    • Due to fewer TLB misses and faster page table look up time

Best Practices

• Install VMware Tools and enable ballooning - better performance than host swapping
• Provide sufficient swap space in guest
• Place guest swap file/partition on separate vdisk
• Don't disable memory compression
• Host cache is "nice to have"

Optimize Host Swapping

• Sharing before swap
• Compressing before swap
• Swap to host cache (SSD)
  • Target 20% for SSD
  • Too big host cache is potentially a waste

Transient memory Pressure

Transient memory pressure is not a problem