2.2.4. Install as a VM using KVM¶

This chapter explains how to start a VM using KVM.

First, you should have a look at the hypervisor prerequisites section.

After the prerequisites are completed, you have two choices:

a simple configuration to try Virtual Service Router CLI using a VM with virtual NICs
a more complex configuration with good performance using a VM with physical NICs

Hypervisor prerequisites¶

We will not detail how to install a linux distribution here. Once it is installed, some tasks must be completed to configure the distribution into an hypervisor.

The kvm and kvm_intel modules have to be inserted:

# lsmod | grep kvm
kvm_intel             172032  0
kvm                   544768  1 kvm_intel

qemu-kvm, libvirt and virt-install have to be installed:

# apt-get install -y qemu-kvm
# apt-get install -y libvirt-daemon-system virtinst libvirt-clients bridge-utils

or

# yum install -y qemu-kvm
# yum install -y libvirt-daemon-system virtinst libvirt-clients bridge-utils

Configuring the halt polling system (halt_poll_ns parameter)¶

The KVM halt polling system is a kernel mechanism that determines how KVM guest vCPUs are handled when they become idle: the VM CPU can be polled for new work from the host during some time. It provides a latency advantage in cases where the guest can be run again quickly, at the price of an increased CPU usage.

The duration of the polling loop is configurable through a KVM module parameter, which is also modifiable at run time:

# echo $TIME_IN_NANOSECONDS > /sys/module/kvm/parameters/halt_poll_ns

Warning

Hypervisor processes (qemu-system-x86_64 for instance) must be restarted after changing this parameter.

We recommend to disable the polling system, to reduce the power consumption when the VM is idle:

# echo 0 > /sys/module/kvm/parameters/halt_poll_ns

This change can be made permanent by adding the option in modprobe configuration:

# echo "options kvm halt_poll_ns=0" > "/etc/modprobe.d/kvm-halt-poll.conf"

VM with virtual NICs¶

In this example, the VM will have three interfaces:

one management interface on the libvirt default virtual network using NAT forwarding,
two data plane interfaces on top of the host’s interfaces using bridged networking to connect the VM to the LAN.

VM with physical NICs¶

This section details how to start Virtual Service Router with dedicated physical NICs.

Using dedicated NICs requires some work which is detailed in Hypervisor mandatory prerequisites.

Once the hypervisor is configured properly, two technologies are available:

whole NICs are dedicated to Virtual Service Router, see Passthrough mode, simpler configuration, but only one VM can use each NIC
portions of NICs are dedicated to Virtual Service Router, see SR-IOV mode, to have more VMs running on the hypervisor

For production setups, you might want to consider checking Optimize performance in virtual environment to get the best performance.

Hypervisor mandatory prerequisites¶

enable Intel VT-d¶

Intel VT-d stands for “Intel Virtualization Technology for Directed I/O”. It is needed to give a physical NIC to a VM. To enable it:

it usually has to be enabled from the BIOS. The name of this feature can differ from one hardware to the other, we advise you to check your hardware documentation to enable it.
it has to be enabled also in the kernel, by adding intel_iommu=on iommu=pt in the kernel command line.

To do so, run:

# echo 'GRUB_CMDLINE_LINUX="$GRUB_CMDLINE_LINUX intel_iommu=on iommu=pt"' \
 >> /etc/default/grub
# update-grub2
# reboot

You can check the boot logs at next boot to verify that Intel VT-d is properly enabled.

# dmesg |grep "Intel(R) Virtualization Technology for Directed I/O"
[    1.391229] DMAR: Intel(R) Virtualization Technology for Directed I/O

hugepages¶

For performance reasons, the memory used by the VMs that will harbor Virtual Service Router must be reserved in hugepages.

Note

A hugepage is a page that addresses more memory than the usual 4KB. Accessing a hugepage is more efficient than accessing a regular memory page. Its default size is 2MB.

hugeadm can be used to managed hugepages. It is part of the hugepages deb package and libhugetlbfs-utils rpm package.

To see if your system already has hugepages available, and which sizes are supported, do:

# hugeadm --pool-list
      Size  Minimum  Current  Maximum  Default
   2097152        0        0        0        *
1073741824        0        0        0

On this system, 2MB and 1GB pages are supported.

If your hardware has several sockets, for performance reason, the memory should be allocated on the same node as the interfaces that will be dedicated to the Virtual Service Router VM.

numactl can show which memory node should be chosen for a particular interface. Look for membind in the following command output. This NIC is on memory node 1.

# numactl -m netdev:ens4f0 --show
policy: bind
preferred node: 1
physcpubind: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
cpubind: 0 1
nodebind: 0 1
membind: 1

Add 8 1GB hugepages for one Virtual Service Router VM to NUMA node 1. You should add this command to a custom startup script to make it persistent.
```
# echo 8 > /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/nr_hugepages
```

Check that the pages were allocated

# hugeadm --pool-list
      Size  Minimum  Current  Maximum  Default
   2097152        0        0        0        *
1073741824        8        8        8

Passthrough mode¶

With this configuration, the Virtual Service Router VM will get dedicated interfaces.

The passthrough mode is only available if the hypervisor’s hardware supports Intel VT-d, and if it is enabled (see enable Intel VT-d).

You must first find the pci id of the interfaces that will be dedicated to the Virtual Service Router VM.

# lspci |grep Ethernet
00.0 Ethernet controller: Intel Corporation Ethernet Connection X552/X557-AT 10GBASE-T
00.1 Ethernet controller: Intel Corporation Ethernet Connection X552/X557-AT 10GBASE-T
00.0 Ethernet controller: Intel Corporation Ethernet 10G 2P X520 Adapter (rev 01)
00.1 Ethernet controller: Intel Corporation Ethernet 10G 2P X520 Adapter (rev 01)
00.0 Ethernet controller: Intel Corporation I350 Gigabit Network Connection (rev 01)
00.1 Ethernet controller: Intel Corporation I350 Gigabit Network Connection (rev 01)

Then use virt-install to spawn the VM, specifying one host-device argument for each device that you want to dedicate. In this example, we dedicate 03:00.0 and 03:00.1.

# cp vsr.qcow2 /var/lib/libvirt/images/vm1.qcow2
# virt-install --name vm1 --vcpus=3,sockets=1,cores=3,threads=1 \
               --os-variant ubuntu22.04 --cpu host --network=default,model=e1000 \
               --ram 8192 --noautoconsole \
               --import --memorybacking hugepages=yes \
               --disk /var/lib/libvirt/images/vm1.qcow2,device=disk,bus=virtio \
               --host-device 03:00.0 --host-device 03:00.1

Connect to the VM:

# virsh console vm1
Connected to domain vm1
Escape character is ^]

  ____        _____ _   _ ____   __     ______  ____
 / /\ \      / /_ _| \ | |  _ \  \ \   / / ___||  _ \
| '_ \ \ /\ / / | ||  \| | | | |  \ \ / /\___ \| |_) |
| (_) \ V  V /  | || |\  | |_| |   \ V /  ___) |  _ <
 \___/ \_/\_/  |___|_| \_|____/     \_/  |____/|_| \_\

vsr login: admin

To get the best performance, the VM CPUs should be associated to physical CPUs. This is called pinning, and is described in CPU pinning.

The next step is to perform your first configuration.

SR-IOV mode¶

SR-IOV enables an Ethernet port to appear as multiple, separate, physical devices called Virtual Functions (VF). You will need compatible hardware, and Intel VT-d configured. The traffic coming from each VF can not be seen by the other VFs. The performance is almost as good as the performance in passthrough mode.

Being able to split an Ethernet port can increase the VM density on the hypervisor compared to passthrough mode.

In this configuration, the Virtual Service Router VM will get Virtual Functions.

First check if the network interface that you want to use supports SR-IOV and how much VFs can be configured. Here we check for eno1 interface.

# lspci -vvv -s $(ethtool -i eno1 | grep bus-info | awk -F': ' '{print $2}') | grep SR-IOV
         Capabilities: [160 v1] Single Root I/O Virtualization (SR-IOV)
# lspci -vvv -s $(ethtool -i eno1 | grep bus-info | awk -F': ' '{print $2}') | grep VFs
             Initial VFs: 64, Total VFs: 64, Number of VFs: 0, Function Dependency Link: 00

Then add VFs, and check that those VFs were created. You should add this command to a custom startup script to make it persistent.

# echo 2 > /sys/class/net/eno1/device/sriov_numvfs
# lspci | grep Ethernet | grep Virtual
03:10.0 Ethernet controller: Intel Corporation Ethernet Connection X552 Virtual Function
03:10.2 Ethernet controller: Intel Corporation Ethernet Connection X552 Virtual Function

You need to set eno1 up so that VFs are properly detected in the guest VM.
```
# ip link set eno1 up
```

Then use virt-install to spawn the VM, specifying one host-device argument for each VF that you want to give. In this example, we give the VF 03:10.0 to Virtual Service Router.

# cp vsr.qcow2 /var/lib/libvirt/images/vm1.qcow2
# virt-install --name vm1 --vcpus=3,sockets=1,cores=3,threads=1 \
               --os-variant ubuntu22.04 --cpu host --network=default,model=e1000 \
               --ram 8192 --noautoconsole --import \
               --memorybacking hugepages=yes \
               --disk /var/lib/libvirt/images/vm1.qcow2,device=disk,bus=virtio \
               --host-device 03:10.0

Connect to the VM:

# virsh console vm1
Connected to domain vm1
Escape character is ^]

  ____        _____ _   _ ____   __     ______  ____
 / /\ \      / /_ _| \ | |  _ \  \ \   / / ___||  _ \
| '_ \ \ /\ / / | ||  \| | | | |  \ \ / /\___ \| |_) |
| (_) \ V  V /  | || |\  | |_| |   \ V /  ___) |  _ <
 \___/ \_/\_/  |___|_| \_|____/     \_/  |____/|_| \_\

vsr login: admin

To get the best performance, the VM CPUs should be associated to physical CPUs. This is called pinning, and is described in CPU pinning.

The next step is to perform your first configuration.

Optimize performance in virtual environment¶

To get good performance, Virtual Service Router needs dedicated resources. It includes:

NICs
CPUs

The first thing to do is to identify the resources that will be dedicated. This can be done in the Identifying hardware resources section.

Then, all the resources must be properly isolated, and configured, see Isolating and configuring hardware resources.

Identifying hardware resources¶

resource inventory¶

Before identifying the resources that will be dedicated to the Virtual Service Router VM, you need to know which NICs and CPUs are available.

It can be done using lstopo, which is part of the hwloc package.

# lstopo -p --merge
Machine (31GB total)
  NUMANode P#0 (16GB)
    Core P#0
      PU P#0
      PU P#20
    Core P#1
      PU P#1
      PU P#21
(...)
    Core P#12
      PU P#9
      PU P#29
    HostBridge P#0
      PCIBridge
        PCI 1000:005b
      PCIBridge
        PCI 15b3:1013
        PCI 15b3:1013
          Net "ens1f1"
      PCIBridge
        PCI 8086:1d6b
      PCIBridge
        PCI 8086:1521
          Net "mgmt0"
        PCI 8086:1521
          Net "enp5s0f1"
        PCI 8086:1521
          Net "enp5s0f2"
        PCI 8086:1521
          Net "enp5s0f3"
      PCIBridge
        PCI 102b:0522
      PCI 8086:1d00
        Block(Disk) "sda"
      PCI 8086:1d08
  NUMANode P#1 (16GB)
    Core P#0
      PU P#10
      PU P#30
    Core P#1
      PU P#11
      PU P#31
(...)
    Core P#12
      PU P#19
      PU P#39
    HostBridge P#2
      PCIBridge
        PCI 8086:1583
        PCI 8086:1583
      PCIBridge
        PCI 8086:1583
          Net "ens4f0"
        PCI 8086:1583

On this machine:

logical CPUs 0 to 9, and ens1f1, mgmt0, enp5s0f1, enp5s0f2, and enp5s0f1 interfaces use NUMA node 0
logical CPUs 10 to 19, and the ens4f0 interface use NUMA node 1

Note

NUMA (Non-uniform memory access) is a memory design, in which a hardware resource can access local memory faster than non-local memory. The memory is organized into several NUMA nodes.

resource dedication¶

Now that you identified your hardware, you can select which NICs and CPUs will be dedicated.

There are some constraints:

we leave the first cpu for Linux
CPUs must be taken on the same node as NICs
crossing NUMA nodes costs performance, so all NICs should be taken on the same node

We recommend to start with a few CPUs, and increase when the setup is functional if needed. The example in this chapter use 3 virtual CPUs.

Isolating and configuring hardware resources¶

CPU isolation¶

The CPUs that will be dedicated to the Virtual Service Router VM need to be properly isolated from other processes. The more reliable way to achieve this is to isolate the CPUs at boot time, on the kernel command line, using the isolcpus and rcu_nocbs directives. For instance, adding isolcpus=1-12,29-40 rcu_nocbs=1-12,29-40 will isolate CPUs 1 to 12 and 29 to 40. It can be added to the kernel command line by doing:

# echo 'GRUB_CMDLINE_LINUX="$GRUB_CMDLINE_LINUX isolcpus=1-12,29-40 rcu_nocbs=1-12,29-40"' >> /etc/default/grub
# update-grub2
# reboot

CPU pinning¶

After the vm is created, you can use virsh vcpupin vm1 vm-cpu cpu to do the one-to-one pinning, using the isolated CPUs. The CPUs should be taken in the list of dedicated CPUs obtained in Identifying hardware resources. The setup is persistent.

For instance, the next commands will pin:

virtual CPU 0 and CPU 2,
virtual CPU 1 and CPU 10,
virtual CPU 2 and CPU 4

# virsh vcpupin vm1 0 2
# virsh vcpupin vm1 1 10
# virsh vcpupin vm1 2 4

CPU configuration¶

The hypervisor CPUs have to be configured for several reasons.

To get stable performance, it is better to disable intel_pstate from the kernel command line:

# echo 'GRUB_CMDLINE_LINUX="$GRUB_CMDLINE_LINUX intel_pstate=disable"' >> /etc/default/grub
# update-grub2
# reboot

To get better performance, the CPUs should use the performance governor. You should add this command to a custom startup script to make it persistent.
```
# cpupower set -b 0
# cpupower frequency-set -g performance
```

For persistent configuration, the previous commands can be added to a custom startup script.

IRQ affinities configuration¶

Having IRQ triggered on the CPUs that are dedicated to the Virtual Service Router VM can result in a few packets lost from time to time. If you don’t notice this problem during testing, you don’t need to take care of this step.

To do so, first ensure that the irqbalance package is removed.

# apt-get remove -y irqbalance

or

# yum remove -y irqbalance

Then run this script:

for file in $(ls /proc/irq)
do
   if [ -f /proc/irq/$file/smp_affinity_list ]; then
      echo "irq: $file"
      echo 0-4,7 > /proc/irq/$file/smp_affinity_list
      mask=$(cat /proc/irq/$file/smp_affinity)
   fi
done
echo $mask > /proc/irq/default_smp_affinity

0-4,7 should be changed to the list of CPUs that are not dedicated to the Virtual Service Router VM.

For persistent configuration, the previous commands can be added to a custom startup script.