This page describes how to test the libvirt driver’s handling of the NUMA placement, large page allocation and CPU pinning features. It relies on setting up a virtual machine as the test environment and requires support for nested virtualization since plain QEMU is not sufficiently functional. The virtual machine will itself be given NUMA topology, so it can then act as a virtual “host” for testing purposes.
The entire test process will take place inside a large virtual machine running Fedora 21. The instructions should work for any other Linux distribution which includes libvirt >= 1.2.9 and QEMU >= 2.1.2
The tests will require support for nested KVM, which is not enabled by default on hypervisor hosts. It must be explicitly turned on in the host when loading the kvm-intel/kvm-amd kernel modules.
On Intel hosts verify it with
# cat /sys/module/kvm_intel/parameters/nested
N
# rmmod kvm-intel
# echo "options kvm-intel nested=y" > /etc/modprobe.d/dist.conf
# modprobe kvm-intel
# cat /sys/module/kvm_intel/parameters/nested
Y
While on AMD hosts verify it with
# cat /sys/module/kvm_amd/parameters/nested
0
# rmmod kvm-amd
# echo "options kvm-amd nested=1" > /etc/modprobe.d/dist.conf
# modprobe kvm-amd
# cat /sys/module/kvm_amd/parameters/nested
1
The virt-install command below shows how to provision a basic Fedora 21 x86_64 guest with 8 virtual CPUs, 8 GB of RAM and 20 GB of disk space:
# cd /var/lib/libvirt/images
# wget http://download.fedoraproject.org/pub/fedora/linux/releases/test/21-Alpha/Server/x86_64/iso/Fedora-Server-netinst-x86_64-21_Alpha.iso
# virt-install \
--name f21x86_64 \
--ram 8000 \
--vcpus 8 \
--file /var/lib/libvirt/images/f21x86_64.img \
--file-size 20
--cdrom /var/lib/libvirt/images/Fedora-Server-netinst-x86_64-21_Alpha.iso \
--os-variant fedora20
When the virt-viewer application displays the installer, follow the defaults for the installation with a couple of exceptions
Once the installation process has completed, the virtual machine will reboot into the final operating system. It is now ready to deploy an OpenStack development environment.
For later ease of use, copy your SSH public key into the virtual machine
# ssh-copy-id <IP of VM>
Now login to the virtual machine
# ssh <IP of VM>
We’ll install devstack under $HOME/src/cloud/.
# mkdir -p $HOME/src/cloud
# cd $HOME/src/cloud
# chmod go+rx $HOME
The Fedora minimal install does not contain git and only has the crude & old-fashioned “vi” editor.
# sudo yum -y install git emacs
At this point a fairly standard devstack setup can be done. The config below is just an example that is convenient to use to place everything in $HOME instead of /opt/stack. Change the IP addresses to something appropriate for your environment of course
# git clone git://github.com/openstack-dev/devstack.git
# cd devstack
# cat >>local.conf <<EOF
[[local|localrc]]
DEST=$HOME/src/cloud
DATA_DIR=$DEST/data
SERVICE_DIR=$DEST/status
LOGFILE=$DATA_DIR/logs/stack.log
SCREEN_LOGDIR=$DATA_DIR/logs
VERBOSE=True
disable_service neutron
HOST_IP=192.168.122.50
FLAT_INTERFACE=eth0
FIXED_RANGE=192.168.128.0/24
FIXED_NETWORK_SIZE=256
FLOATING_RANGE=192.168.129.0/24
MYSQL_PASSWORD=123456
SERVICE_TOKEN=123456
SERVICE_PASSWORD=123456
ADMIN_PASSWORD=123456
RABBIT_PASSWORD=123456
IMAGE_URLS="http://download.cirros-cloud.net/0.3.2/cirros-0.3.2-x86_64-uec.tar.gz"
EOF
# FORCE=yes ./stack.sh
Unfortunately while devstack starts various system services and changes various system settings it doesn’t make the changes persistent. Fix that now to avoid later surprises after reboots
# sudo systemctl enable mysqld.service
# sudo systemctl enable rabbitmq-server.service
# sudo systemctl enable httpd.service
# sudo emacs /etc/sysconfig/selinux
SELINUX=permissive
First to confirm we’ve not done anything unusual to the traditional operation of Nova libvirt guests boot a tiny instance
# . openrc admin
# nova boot --image cirros-0.3.2-x86_64-uec --flavor m1.tiny cirros1
The host will be reporting NUMA topology, but there should only be a single NUMA cell this point.
# mysql -u root -p nova
MariaDB [nova]> select numa_topology from compute_nodes;
+----------------------------------------------------------------------------+
| numa_topology |
+----------------------------------------------------------------------------+
| {
| "nova_object.name": "NUMATopology",
| "nova_object.data": {
| "cells": [{
| "nova_object.name": "NUMACell",
| "nova_object.data": {
| "cpu_usage": 0,
| "memory_usage": 0,
| "cpuset": [0, 1, 2, 3, 4, 5, 6, 7],
| "pinned_cpus": [],
| "siblings": [],
| "memory": 7793,
| "mempages": [
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 987430,
| "used": 0,
| "size_kb": 4
| },
| },
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 0,
| "used": 0,
| "size_kb": 2048
| },
| }
| ],
| "id": 0
| },
| },
| ]
| },
| }
+----------------------------------------------------------------------------+
Meanwhile, the guest instance should not have any NUMA configuration recorded
MariaDB [nova]> select numa_topology from instance_extra;
+---------------+
| numa_topology |
+---------------+
| NULL |
+---------------+
Now that devstack is proved operational, it is time to configure some NUMA topology for the test VM, so that it can be used to verify the OpenStack NUMA support. To do the changes, the VM instance that is running devstack must be shut down.
# sudo shutdown -h now
And now back on the physical host edit the guest config as root
# sudo virsh edit f21x86_64
The first thing is to change the <cpu> block to do passthrough of the host CPU. In particular this exposes the “SVM” or “VMX” feature bits to the guest so that “Nested KVM” can work. At the same time we want to define the NUMA topology of the guest. To make things interesting we’re going to give the guest an asymmetric topology with 4 CPUS and 4 GBs of RAM in the first NUMA node and 2 CPUs and 2 GB of RAM in the second and third NUMA nodes. So modify the guest XML to include the following CPU XML
<cpu mode='host-passthrough'>
<numa>
<cell id='0' cpus='0-3' memory='4096000'/>
<cell id='1' cpus='4-5' memory='2048000'/>
<cell id='2' cpus='6-7' memory='2048000'/>
</numa>
</cpu>
The guest can now be started again, and ssh back into it
# virsh start f21x86_64
...wait for it to finish booting
# ssh <IP of VM>
Before starting OpenStack services again, it is necessary to reconfigure Nova to enable the NUMA scheduler filter. The libvirt virtualization type must also be explicitly set to KVM, so that guests can take advantage of nested KVM.
# sudo emacs /etc/nova/nova.conf
Set the following parameters:
[DEFAULT]
scheduler_default_filters=RetryFilter, AvailabilityZoneFilter, RamFilter, ComputeFilter, ComputeCapabilitiesFilter, ImagePropertiesFilter, ServerGroupAntiAffinityFilter, ServerGroupAffinityFilter, NUMATopologyFilter
[libvirt]
virt_type = kvm
With that done, OpenStack can be started again
# cd $HOME/src/cloud/devstack
# ./rejoin-stack.sh
The first thing is to check that the compute node picked up the new NUMA topology setup for the guest
# mysql -u root -p nova
MariaDB [nova]> select numa_topology from compute_nodes;
+----------------------------------------------------------------------------+
| numa_topology |
+----------------------------------------------------------------------------+
| {
| "nova_object.name": "NUMATopology",
| "nova_object.data": {
| "cells": [{
| "nova_object.name": "NUMACell",
| "nova_object.data": {
| "cpu_usage": 0,
| "memory_usage": 0,
| "cpuset": [0, 1, 2, 3],
| "pinned_cpus": [],
| "siblings": [],
| "memory": 3857,
| "mempages": [
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 987430,
| "used": 0,
| "size_kb": 4
| },
| },
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 0,
| "used": 0,
| "size_kb": 2048
| },
| }
| ],
| "id": 0
| },
| },
| {
| "nova_object.name": "NUMACell",
| "nova_object.data": {
| "cpu_usage": 0,
| "memory_usage": 0,
| "cpuset": [4, 5],
| "pinned_cpus": [],
| "siblings": [],
| "memory": 1969,
| "mempages": [
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 504216,
| "used": 0,
| "size_kb": 4
| },
| },
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 0,
| "used": 0,
| "size_kb": 2048
| },
| }
| ],
| "id": 1
| },
| },
| {
| "nova_object.name": "NUMACell",
| "nova_object.data": {
| "cpu_usage": 0,
| "memory_usage": 0,
| "cpuset": [6, 7],
| "pinned_cpus": [],
| "siblings": [],
| "memory": 1967,
| "mempages": [
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 503575,
| "used": 0,
| "size_kb": 4
| },
| },
| {
| "nova_object.name": "NUMAPagesTopology",
| "nova_object.data": {
| "total": 0,
| "used": 0,
| "size_kb": 2048
| },
| }
| ],
| "id": 2
| },
| }
| ]
| },
| }
+----------------------------------------------------------------------------+
This indeed shows that there are now 3 NUMA nodes for the “host” machine, the first with 4 GB of RAM and 4 CPUs, and others with 2 GB of RAM and 2 CPUs each.
For the sake of backwards compatibility, if the NUMA filter is enabled, but the flavor/image does not have any NUMA settings requested, it should be assumed that the guest will have a single NUMA node. The guest should be locked to a single host NUMA node too. Boot a guest with the m1.tiny flavor to test this condition
# . openrc admin admin
# nova boot --image cirros-0.3.2-x86_64-uec --flavor m1.tiny cirros1
Now look at the libvirt guest XML. It should show that the vCPUs are locked to pCPUs within a particular node.
# virsh -c qemu:///system list
....
# virsh -c qemu:///system dumpxml instanceXXXXXX
...
<vcpu placement='static' cpuset='6-7'>1</vcpu>
...
This example shows that the guest has been locked to the 3rd NUMA node (which contains pCPUs 6 and 7). Note that there is no explicit NUMA topology listed in the guest XML.
Moving forward a little, explicitly tell Nova that the NUMA topology for the guest should have a single NUMA node. This should operate in an identical manner to the default behavior where no NUMA policy is set. To define the topology we will create a new flavor
# nova flavor-create m1.numa 999 1024 1 4
# nova flavor-key m1.numa set hw:numa_nodes=1
# nova flavor-show m1.numa
Now boot the guest using this new flavor
# nova boot --image cirros-0.3.2-x86_64-uec --flavor m1.numa cirros2
Looking at the resulting guest XML from libvirt
# virsh -c qemu:///system dumpxml instanceXXXXXX
...
<vcpu placement='static'>4</vcpu>
<cputune>
<vcpupin vcpu='0' cpuset='0-3'/>
<vcpupin vcpu='1' cpuset='0-3'/>
<vcpupin vcpu='2' cpuset='0-3'/>
<vcpupin vcpu='3' cpuset='0-3'/>
<emulatorpin cpuset='0-3'/>
</cputune>
...
<cpu>
<topology sockets='4' cores='1' threads='1'/>
<numa>
<cell id='0' cpus='0-3' memory='1048576'/>
</numa>
</cpu>
...
<numatune>
<memory mode='strict' nodeset='0'/>
<memnode cellid='0' mode='strict' nodeset='0'/>
</numatune>
The XML shows:
As a further sanity test, check what Nova recorded for the instance in the database. This should match the <numatune> information
MariaDB [nova]> select numa_topology from instance_extra;
+----------------------------------------------------------------------------+
| numa_topology |
+----------------------------------------------------------------------------+
| {
| "nova_object.name": "InstanceNUMATopology",
| "nova_object.data": {
| "instance_uuid": "4c2302fe-3f0f-46f1-9f3e-244011f6e03a",
| "cells": [
| {
| "nova_object.name": "InstanceNUMACell",
| "nova_object.data": {
| "cpu_topology": null,
| "pagesize": null,
| "cpuset": [
| 0,
| 1,
| 2,
| 3
| ],
| "memory": 1024,
| "cpu_pinning_raw": null,
| "id": 0
| },
| }
| ]
| },
| }
+----------------------------------------------------------------------------+
Now getting more advanced we tell Nova that the guest will have two NUMA nodes. To define the topology we will change the previously defined flavor
# nova flavor-key m1.numa set hw:numa_nodes=2
# nova flavor-show m1.numa
Now boot the guest using this changed flavor
# nova boot --image cirros-0.3.2-x86_64-uec --flavor m1.numa cirros2
Looking at the resulting guest XML from libvirt
# virsh -c qemu:///system dumpxml instanceXXXXXX
...
<vcpu placement='static'>4</vcpu>
<cputune>
<vcpupin vcpu='0' cpuset='0-3'/>
<vcpupin vcpu='1' cpuset='0-3'/>
<vcpupin vcpu='2' cpuset='4-5'/>
<vcpupin vcpu='3' cpuset='4-5'/>
<emulatorpin cpuset='0-5'/>
</cputune>
...
<cpu>
<topology sockets='4' cores='1' threads='1'/>
<numa>
<cell id='0' cpus='0-1' memory='524288'/>
<cell id='1' cpus='2-3' memory='524288'/>
</numa>
</cpu>
...
<numatune>
<memory mode='strict' nodeset='0-1'/>
<memnode cellid='0' mode='strict' nodeset='0'/>
<memnode cellid='1' mode='strict' nodeset='1'/>
</numatune>
The XML shows:
As a further sanity test, check what Nova recorded for the instance in the database. This should match the <numatune> information
MariaDB [nova]> select numa_topology from instance_extra;
+----------------------------------------------------------------------------+
| numa_topology |
+----------------------------------------------------------------------------+
| {
| "nova_object.name": "InstanceNUMATopology",
| "nova_object.data": {
| "instance_uuid": "a14fcd68-567e-4d71-aaa4-a12f23f16d14",
| "cells": [
| {
| "nova_object.name": "InstanceNUMACell",
| "nova_object.data": {
| "cpu_topology": null,
| "pagesize": null,
| "cpuset": [
| 0,
| 1
| ],
| "memory": 512,
| "cpu_pinning_raw": null,
| "id": 0
| },
| },
| {
| "nova_object.name": "InstanceNUMACell",
| "nova_object.data": {
| "cpu_topology": null,
| "pagesize": null,
| "cpuset": [
| 2,
| 3
| ],
| "memory": 512,
| "cpu_pinning_raw": null,
| "id": 1
| },
| }
| ]
| },
| }
|
+----------------------------------------------------------------------------+
Enter search terms or a module, class or function name.