More info on http://techdays.be.

1. Hyper-V Dynamic Memory in Depth

2. Agenda
• Memory & Virtualization
• Memory Optimization Techniques
• Hyper-V Dynamic Memory
• Architecture & Concepts
• Demo
• System Impact
• Q&A

4. Virtualization & Memory
• How much memory does a server actually need?
• IIS Server?
• Print server?
• File server?
• Branch Cache?
• Direct Access?
• How much will performance be affected if you halved the
amount of memory in a VM?

5. No one wants to size VMs Memory
• “New virtual machines get 1GB of RAM [no matter what the VM is running]. I
only give people more memory if they complain about performance”
• “All VMs get 4GB of RAM [I have no idea what is happening with that memory]
and no one complains”
• “I take the minimum system requirements and add (insert one: 50%, 100%,
150%)”
• “A vendor tells me their app needs 4GB of RAM. I do not have the time to test
this to find out if it is true or not”

6. Virtualization & Memory
• Memory
• Key Factor to the number of running VMs
• The most utilized asset in the system, but also a more expensive asset
• Statistics on resource utilization of workloads
• CPU 10%
• Memory 40%
• Network I/O <5%
• Disk I/O <5%
• Customer Requirements
• Maximum density, without sacrificing performance
• Maintain consistent performance
• Don’t provide a feature that’s unsuitable for production use

7. Dynamic Memory Goals
• Higher VM consolidation ratios with minimal performance
impact
• Dependent on:
• How much variation in memory utilization the workloads have
• How good a job you did of sizing the systems in the first place
• Work well for both server and desktop workloads
• Add minimal overhead to the system
• Pass the “that looks right” test

8. Understanding philosophical differences
• Host and Guest Collaborate
• Understands what guest information to use
• Trying to get the “best bang for buck” in virtual memory
management
• Host works in Isolation
• Does not trust guest information
• Building a “black box” solution

9. Dynamic Memory, not Overcommit
• Overloaded Term
• Over-subscription vs. Over-committment
• No one wants to overcommit their resource
• You don’t overcommit other resources (really – you do not)
• VMware does not want you to overcommit memory (really)
• Dynamic Memory treat memory like we treat CPU
resources
• Dynamically schedulable resource

11. Ballooning
• How it works:
• Increasing the size of the balloon forces the guest to react to memory pressure by
releasing unused pages
• Decreasing the size makes more memory available to the guest
Swap Out Swap In
Memory Memory Memory
Balloon Balloon Balloon
Disk Disk Ballloon Disk
Inflate Deflate
Deflate

12. Second Level Paging
• How it works:
• Paging at the virtualization layer by swapping VMs
memory out to disk
• Many problems:
• Swapping Guest Kernel Resources
• Double Paging
• Disks are slow
• But it always works…

13. External Page Sharing
• How it works:
• Eliminate redundant copies of memory pages common to more than
one virtual machine
1. Hash all memory and store it in a table…
2. Identify the common hashes and then…
3. Perform a bit by bit comparison
• Problems
• Page Sharing not dynamic
• Can take hours to share pages
• The largest benefit are zero pages
• Doesn’t work with large pages

14. Other Techniques
• Guest directed page sharing
• Memory compression
• And on…

16. System Requirements
• Parent Requirements:
• Windows Server 2008 R2 SP1, Windows Server 2012
• Hyper-V Server 2008 R2 SP1, Hyper-V Server 2012
• Guest Requirements:
• Windows Server 2003, 2008 & 2008 R2, 2012
• 32-bit & 64-bit versions
• Windows Vista, Windows 7, Windows 8
• Enterprise and Ultimate Editions only
• 32-bit & 64-bit versions

17. Adding/Removing Memory
•Adding Memory
• Enlightened fashion
• Synthetic Memory Driver (VSP/VSC Pair)
• No hardware emulation
• Light weight
•Removing Memory
• Wanted to remove memory
• Ballooning is more efficient
• Messes up task manager in the guest OS

18. Adding/Removing Memory
• Active Memory addition
• Memory is added immediately when VM needs it
• Passive Memory reclamation
• Memory is not removed when there is no immediate need
• Unutilized memory is collected every 5 minutes

19. Startup & Max
Windows Server 2008 R2 SP1
• Startup: amount of memory to boot VM
• BIOS does not know about DM
• Guest OS may not know about DM
• Default: 512MB
• Max: don’t let the VM above this amount
• Default: 64GB
Windows Server 2012
• Startup: amount of memory to boot VM
• Min: amount of memory the VM can balloon down to
• Max: don’t let the VM above this amount

20. Controlling Memory Availability
• Availability is a concept
• How much memory does the VM have?
• How much memory does the VM want?
• The difference is the availability
• Mechanisms in place to control Memory Availability
through Buffer and Priority settings

21. Memory Buffer & Priority
• Buffer: How much “free” memory should we try and keep in the
VM?
• Allows for responsiveness to bursty workloads
• Can be used for file cache
“I like to configure my virtual machines so that they have ~20% free
memory”
• Priority: which VM gets the memory first
• 1-10,000: default is 5,000
• The higher the priority, the higher the availability

24. Changes to Root Reserve
• Hyper-V has always had the concept of a reserve of
memory that is kept for the parent partition
VM
Host Reserve Overhead Memory utilized by VMs
VM Reserve

25. Changes to Root Reserve
• DM allows VMs to push up against the reserve
consistently
• New behavior to better protect the parent partition
from rampaging virtual machines
• New registry key in place
• Allows you to reserve static memory for the parent partition
• May result in less memory being available for VMs

26. Changes to NUMA management…
• Wait – what is NUMA?
• Why do I care?
• How does this work today?

27. What is “NUMA”?
• A traditional computer:

28. What is “NUMA”?
• A NUMA computer:

29. Why do I care?
• VM memory should come from the “local NUMA node”

30. Why do I care?
• Ideally VM memory should come from the “local NUMA node”

31. How does this work today?
• Hyper-V tries to get all memory for a virtual machine from a single
NUMA node
• When it cannot – the virtual machine “spans” NUMA nodes
• Users can set preferred NUMA nodes for virtual machines in order
to get the best distribution

32. Changes to NUMA management
• Dynamic memory can result in more virtual machines
spanning NUMA nodes
• A virtual machine might start all on one node – but added
memory might come from another node
• New option to disable NUMA node spanning

33. Disabling NUMA Spanning
• Makes the system behave like multiple small computers