NUMA aware placement for KVM

Description This issue is to implement placement constraints on NUMA nodes. A VM can be pinned on a node based on:

• PCI affinity
• Memory affinity
• Manual node assigment

Use case Improve performance of VM by making memory and/or I/O access to the same NUMA node.

Interface Changes New atrributes to set NUMA placement preferences:

NUMA = [
   NODE_POLICY = <PCI | MEMORY | AUTO | MANUAL >
   ALLOCATION_POLICY = <DEDICATED | NODE >
   NUMA_NODE_ID = <>
   HUGE_PAGE_SIZE_MB = <>
]

NUMA node selection (NODE_POLICY attribute):

• PCI affinity. Place the VM in the same numa node as the PCI passthrough devices. If multiple devices are present the VM will be placed in the node of the first one.
• MEMORY guided. Place the VM in the first numa node with enough free memory
• MANUAL place the VM in the given node
• AUTO: : As assgined by numad implies (default):

<numatune>
        <memory mode='strict' placement='auto'/>
</numatune>
...
<vcpu placement='auto'>2</vcpu>

vCPU and Memory will be assigned according to ALLOCATION_POLICY as follows:

• NODE: OpenNebula will pin the vCPUs an memory to the selected NUMA core resources. (default)

<vcpu placement='static' cpuset='0-13,28-41'>2</vcpu>
...
<numatune>
        <memory mode='strict' nodeset='0'/>
</numatune>

Other VMs can be assigned to the same NUMA cores in the node.

• DEDICATED: Same as node but each vCPU is pinned to an specific NUMA core. OpenNebula will also mark this node as used and will not be assigned to any other VM.

<cputune> 
  <vcpupin vcpu='0' cpuset='0'/> 
  <vcpupin vcpu='1' cpuset='2'/> 
 </cputune> 

Additional Context N/A

Progress Status

• Branch created
• Code committed to development branch
• Testing - QA
• Documentation
• Release notes - resolved issues, compatibility, known issues
• Code committed to upstream release/hotfix branches
• Documentation committed to upstream release/hotfix branches

[paczerny]

Hi, I’m currently working on #3673. Not only CPU_MODEL, but all configuration options from vmm_exec_kvm.conf will be possible to specify in cluster or in the host. The priority of the values will be VM > Host > Cluster > config file

But the RAW configuration option works slightly different now. The RAW from the config file is added to the VM RAW. This behavior must preserve. So the logic for RAW will be: use VM RAW and add first RAW option from Host, Cluster or config file.

[feldsam]

Hi, thanks for reply. I think that my request is more related to this #3664 because what I doing in raw xml, is de facto pinning to numa node and enabling hugepages.