Cilium
Versions Supported
- 1.15.x
- 1.14.x
If you are using Bring Your Own Operating System (BYOOS), then HWE (Hardware Enabled) Kernel or a Kernel that supports eBPF modules needs to be provisioned.
Disable UDP Segmentation on VMXNET3 Edge Hosts for Clusters with Overlay Networks
Due to a known issue with VMware's VMXNET3 adapter, which is widely used in different virtual machine management services, including VMware vSphere and Hyper-V, Cilium Pods may face network connectivity issues.
If deploying an Edge host in a virtual machine environment using a VMXNET3 adapter and enabling an overlay network for your cluster, add the following commands in the user-data
file at the boot stage. Replace <interface-name>
with the name of the network interface on your Edge host.
stages:
initramfs:
- name: "Disable UDP segmentation"
commands:
- ethtool --offload <interface-name> tx-udp_tnl-segmentation off
- ethtool --offload <interface-name> tx-udp_tnl-csum-segmentation off
Troubleshooting
Scenario - I/O Timeout Error on VMware
If you are deploying a cluster to a VMware environment using the VXLAN tunnel protocol, you may encounter I/O timeout errors. This is due to a known bug in the VXMNET3 adapter that results in VXLAN traffic being dropped. You can learn more about this issue in Cilium's GitHub issue #21801.
You can work around the issue by using one of the two following methods:
-
Option 1: Set a different tunnel protocol in the Cilium configuration. You can set the tunnel protocol to
geneve
.charts:
cilium:
tunnelProtocol: "geneve" -
Option 2: Modify the Operating System (OS) layer of your cluster profile to automatically disable UDP Segmentation Offloading (USO).
kubeadmconfig:
preKubeadmCommands:
# Disable hardware segmentation offloading due to VMXNET3 issue
- |
install -m 0755 /dev/null /usr/lib/networkd-dispatcher/routable.d/10-disable-offloading
cat <<EOF > /usr/lib/networkd-dispatcher/routable.d/10-disable-offloading
#!/bin/sh
ethtool -K eth0 tx-udp_tnl-segmentation off
ethtool -K eth0 tx-udp_tnl-csum-segmentation off
ethtool --offload eth0 rx off tx off
EOF
systemctl restart systemd-networkd
Prerequisite
If you are using Bring Your Own Operating System (BYOOS), then HWE (Hardware Enabled) Kernel or a Kernel that supports eBPF modules needs to be provisioned.
Troubleshooting
Scenario - I/O Timeout Error on VMware
If you are deploying a cluster to a VMware environment using the VXLAN tunnel protocol, you may encounter I/O timeout errors. This is due to a known bug in the VXMNET3 adapter that results in VXLAN traffic being dropped. You can learn more about this issue in Cilium's GitHub issue #21801.
You can work around the issue by using one of the two following methods:
-
Option 1: Set a different tunnel protocol in the Cilium configuration. You can set the tunnel protocol to
geneve
.charts:
cilium:
tunnelProtocol: "geneve" -
Option 2: Modify the Operating System (OS) layer of your cluster profile to automatically disable UDP Segmentation Offloading (USO).
kubeadmconfig:
preKubeadmCommands:
# Disable hardware segmentation offloading due to VMXNET3 issue
- |
install -m 0755 /dev/null /usr/lib/networkd-dispatcher/routable.d/10-disable-offloading
cat <<EOF > /usr/lib/networkd-dispatcher/routable.d/10-disable-offloading
#!/bin/sh
ethtool -K eth0 tx-udp_tnl-segmentation off
ethtool -K eth0 tx-udp_tnl-csum-segmentation off
ethtool --offload eth0 rx off tx off
EOF
systemctl restart systemd-networkd
Terraform
You can reference the Cilium pack in Terraform with the following data resource.
data "spectrocloud_registry" "public_registry" {
name = "Public Repo"
}
data "spectrocloud_pack" "cilium" {
name = "cni-cilium-oss"
version = "1.15.3"
registry_uid = data.spectrocloud_registry.public_registry.id
}