Using DPDK
- Without hugepages

Primary interface configuration

the interface is consumed with a native VPP driver. Performance is good and set up is simple, but only virtio interfaces are supported
avf we create a virtual function and consume it with a native VPP driver. Performance is good and setup simple, but only intel AVF interfaces are supported
af_packet the interface stays in Linux which passes packets to VPP. Performance is low, but it works out of the box with any interface
af_xdp packets are passed via eBPF. This requires a >=5.4 kernel, but works out of the box with good performance
dpdk the interface is removed from Linux and consumed with the dpdk library. Performance and support are good, but setup can be complex
other native VPP drivers bring better performance than dpdk but require complex manual setup

The main interface configuration is controlled by a variable named CALICOVPP_NATIVE_DRIVER that is passed to the vpp container. You can edit the .yaml file as follows or use the overlays provided in ./yaml/overlays/*.

By default, leaving CALICOVPP_NATIVE_DRIVER empty (or unspecified) will try all drivers supported in your setup, starting with the most performant. You’ll still need to allocate hugepages if you want e.g. Virtio to work.

For most deployments (except for DPDK) you should ensure that the vpp configuration stanza has DPDK disabled

vpp_config_template: |-
  ...
  # removed dpdk { ... }
  ...
  plugins {
      plugin default { enable }
      plugin calico_plugin.so { enable }
      plugin dpdk_plugin.so { disable }
  }

Using the native Virtio driver

You can use this driver if your primary interface is virtio [realpath /sys/bus/pci/devices/<PCI_ID>/driver gives .../virtio-net]

Ensure you have hugepages available on your system (sysctl -w vm.nr_hugepages=256)
Ensure vfio-pci is loaded (sudo modprobe vfio-pci)
Pass CALICOVPP_NATIVE_DRIVER = virtio to the vpp container

You can use this driver if your primary interface is supported by AVF [realpath /sys/bus/pci/devices/<PCI_ID>/driver gives .../i40e]

Ensure vfio-pci is loaded (sudo modprobe vfio-pci)
Pass CALICOVPP_NATIVE_DRIVER = avf to the vpp container
Also ensure that your vpp config has no dpdk stanza and its plugin disabled
Optionally CALICOVPP_RX_QUEUES controls the number of queues used, CALICOVPP_RING_SIZE their size

Using AF_XDP

caution

Pass CALICOVPP_NATIVE_DRIVER = af_xdp to the vpp container
Also ensure that your vpp config has no stanza and its plugin disabled
Optionally CALICOVPP_RX_QUEUES controls the number of queues used, CALICOVPP_RING_SIZE their size
CALICOVPP_RX_MODE controls whether we busy-poll the interface (polling), only use interrupts to wake us up (interrupt) or switch between both depending on the load (adaptive)
Finally FELIX_XDPENABLED should be set to false on the calico-node container otherwise felix will periodically cleanup the VPP configuration

With kustomize use kubectl kustomize ./yaml/overlays/af-xdp | kubectl apply -f -

Side nodes

AF_XDP won’t start if you specify buffers { buffers-per-numa } to be too big (65536 should work)

Pass CALICOVPP_NATIVE_DRIVER = af_packet to the vpp container
Also ensure that your vpp config has no dpdk stanza and the dpdk plugin is disabled

You can also use kubectl kustomize ./yaml/overlays/af-packet | kubectl apply -f -

Using DPDK

Ensure you have hugepages available on your system (sysctl -w vm.nr_hugepages=256)
Pass CALICOVPP_NATIVE_DRIVER = none to the vpp container
The vpp config in the calico-config ConfigMap should look like the following, __PCI_DEVICE_ID__ will be automatically populated with the PCI ID of CALICOVPP_INTERFACE at startup.
CALICOVPP_RX_QUEUES and CALICOVPP_RING_SIZE have no more effect. They are controlled by their counterparts in the dpdk {} stanza namely num-rx-queues and num-rx-desc

vpp_config_template: |-
  unix {
    nodaemon
    full-coredump
    cli-listen /var/run/vpp/cli.sock
  }
  api-trace { on }
  cpu {
      main-core 1
      workers 0
  }
  socksvr {
      socket-name /var/run/vpp/vpp-api.sock
  dpdk {
    dev __PCI_DEVICE_ID__ { num-rx-queues 1  num-tx-queues 1 }
  plugins {
      plugin default { enable }
      plugin calico_plugin.so { enable }
  }

Without hugepages

DPDK can also run without hugepages with the a configuration similar to the previous one

Pass CALICOVPP_NATIVE_DRIVER = none to the vpp container
The vpp config in the calico-config ConfigMap should look like the following, with __PCI_DEVICE_ID__ automatically populated with that of CALICOVPP_INTERFACE at startup
CALICOVPP_RX_QUEUES and CALICOVPP_RING_SIZE have no more effect. They are controlled by their counterparts in the dpdk {} stanza namely num-rx-queues and num-rx-desc

You also need to tell dpdk to no try allocating hugepages on its own

dpdk { no-hugetlb iova-mode va } does this for the dpdk plugin
buffers { no-hugetlb } does this for the buffers backing VPP’s packets

caution

This won’t run with unsafe_iommu on. You need to echo N | sudo tee /sys/module/vfio/parameters/enable_unsafe_noiommu_mode

This is a rather advanced/experimental setup, we’ll take the example of the AVF driver for this, using vpp cli, but any vpp driver can be used. This allow to efficiently support other interface types.

Pass CALICOVPP_NATIVE_DRIVER = none to the vpp container
Same as before, you should remove the dpdk { ... } section in vpp_config_template and update the plugins { ... } definitions as follows
Also add a exec /etc/vpp/startup.exec entry in unix { .. }

vpp_config_template: |-
  unix {
    nodaemon
    full-coredump
    log /var/run/vpp/vpp.log
    cli-listen /var/run/vpp/cli.sock
    exec /etc/vpp/startup.exec
  }
  ...
  # removed dpdk { ... }
  ...
  plugins {
      plugin default { enable }
      plugin calico_plugin.so { enable }
  }

In the specific case of the AVF driver, the PCI id must belong to a VF that can be created with the . Different drivers will have different requirements.