Skip to content
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
386 lines (245 sloc) 15.2 KB


Networking plugin for pod networking in Kubernetes using Elastic Network Interfaces on AWS.

BuildStatus Widget GoReport Widget


Download the latest version of the yaml and apply it the cluster.

kubectl apply -f aws-k8s-cni.yaml

Launch kubelet with network plugins set to cni (--network-plugin=cni), the cni directories configured (--cni-config-dir and --cni-bin-dir) and node ip set to the primary IPv4 address of the primary ENI for the instance (--node-ip=$(curl It is also recommended to set --max-pods equal to (the number of ENIs for the instance type × (the number of IPs per ENI - 1)) + 2; for details, see vpc_ip_resource_limit.go. Setting --max-pods will prevent scheduling that exceeds the IP address resources available to the kubelet.

The default manifest expects --cni-conf-dir=/etc/cni/net.d and --cni-bin-dir=/opt/cni/bin.

L-IPAM requires following IAM policy:

     "Effect": "Allow",
     "Action": [
     "Resource": [
     "Effect": "Allow",
     "Action": "ec2:CreateTags",
     "Resource": "arn:aws:ec2:*:*:network-interface/*"

Alternatively there is also a Helm chart: eks/aws-vpc-cni


  • make defaults to make build-linux that builds the Linux binaries.
  • unit-test, lint and vet provide ways to run the respective tests/tools and should be run before submitting a PR.
  • make docker will create a docker container using the docker-build with the finished binaries, with a tag of amazon/amazon-k8s-cni:latest
  • make docker-build uses a docker container (golang:1.12) to build the binaries.
  • make docker-unit-tests uses a docker container (golang:1.12) to run all unit tests.


There are 2 components:

  • CNI Plugin, which will wire up host's and pod's network stack when called.
  • L-IPAMD, which is a long running node-Local IP Address Management (IPAM) daemon, is responsible for:
    • maintaining a warm-pool of available IP addresses, and
    • assigning an IP address to a Pod.

The details can be found in Proposal: CNI plugin for Kubernetes networking over AWS VPC.

Troubleshooting Guide provides tips on how to debug and troubleshoot CNI.

ENI Allocation

When a worker node first joins the cluster, there is only 1 ENI along with all of its addresses in the ENI. Without any configuration, ipamD always try to keep one extra ENI.

When number of pods running on the node exceeds the number of addresses on a single ENI, the CNI backend start allocating a new ENI and start using following allocation scheme:

For example, a m4.4xlarge node can have up to 8 ENIs, and each ENI can have up to 30 IP addresses. (

  • If the number of current running Pods is between 0 to 29, ipamD will allocate one more eni. And Warm-Pool size is 2 eni * (30 -1) = 58
  • If the number of current running Pods is between 30 and 58, ipamD will allocate 2 more eni. And Warm-Pool size is 3 eni * (30 -1) = 87

CNI Configuration Variables

The Amazon VPC CNI plugin for Kubernetes supports a number of configuration options, which are set through environment variables. The following environment variables are available, and all of them are optional.


Type: Boolean

Default: true

Specifies whether NodePort services are enabled on a worker node's primary network interface. This requires additional iptables rules and that the kernel's reverse path filter on the primary interface is set to loose.


Type: Boolean

Default: false

Specifies that your pods may use subnets and security groups that are independent of your worker node's VPC configuration. By default, pods share the same subnet and security groups as the worker node's primary interface. Setting this variable to true causes ipamD to use the security groups and VPC subnet in a worker node's ENIConfig for elastic network interface allocation. You must create an ENIConfig custom resource for each subnet that your pods will reside in, and then annotate or label each worker node to use a specific ENIConfig (multiple worker nodes can be annotated or labelled with the same ENIConfig). Worker nodes can only be annotated with a single ENIConfig at a time, and the subnet in the ENIConfig must belong to the same Availability Zone that the worker node resides in. For more information, see CNI Custom Networking in the Amazon EKS User Guide.


Type: String


Specifies node annotation key name. This should be used when AWS_VPC_K8S_CNI_CUSTOM_NETWORK_CFG=true. Annotation value will be used to set ENIConfig name. Note that annotations take precedence over labels.


Type: String


Specifies node label key name. This should be used when AWS_VPC_K8S_CNI_CUSTOM_NETWORK_CFG=true. Label value will be used to set ENIConfig name. Note that annotations will take precedence over labels. To use labels ensure annotation with key or defined key (in ENI_CONFIG_ANNOTATION_DEF) is not set on the node. To select an ENIConfig based upon availability zone set this to and create an ENIConfig custom resource for each availability zone (e.g. us-east-1a).


Type: Integer

Default: 9001

Used to configure the MTU size for attached ENIs. The valid range is from 576 to 9001.


Type: Boolean

Default: false

Specifies whether an external NAT gateway should be used to provide SNAT of secondary ENI IP addresses. If set to true, the SNAT iptables rule and off-VPC IP rule are not applied, and these rules are removed if they have already been applied. Disable SNAT if you need to allow inbound communication to your pods from external VPNs, direct connections, and external VPCs, and your pods do not need to access the Internet directly via an Internet Gateway. However, your nodes must be running in a private subnet and connected to the internet through an AWS NAT Gateway or another external NAT device.


Type: String

Default: hashrandom

Valid Values: hashrandom, prng, none

Specifies weather the SNAT iptables rule should randomize the outgoing ports for connections. This should be used when AWS_VPC_K8S_CNI_EXTERNALSNAT=false. When enabled (hashrandom) the --random flag will be added to the SNAT iptables rule. To use pseudo random number generation rather than hash based (i.e. --random-fully) use prng for the environment variable. For old versions of iptables that do not support --random-fully this option will fall back to --random. Disable (none) this functionality if you rely on sequential port allocation for outgoing connections.

Note: Any options other than none will cause outbound connections to be assigned a source port that's not necessarily part of the ephemeral port range set at the OS level (/proc/sys/net/ipv4/ip_local_port_range). This is relevant for any customers that might have NACLs restricting traffic based on the port range found in ip_local_port_range


Type: String

Default: empty

Specify a comma separated list of IPv4 CIDRs to exclude from SNAT. For every item in the list an iptables rule and off-VPC IP rule will be applied. If an item is not a valid ipv4 range it will be skipped. This should be used when AWS_VPC_K8S_CNI_EXTERNALSNAT=false.


Type: Integer

Default: 1

Specifies the number of free elastic network interfaces (and all of their available IP addresses) that the ipamD daemon should attempt to keep available for pod assignment on the node. By default, ipamD attempts to keep 1 elastic network interface and all of its IP addresses available for pod assignment. The number of IP addresses per network interface varies by instance type. For more information, see IP Addresses Per Network Interface Per Instance Type in the Amazon EC2 User Guide for Linux Instances.

For example, an m4.4xlarge launches with 1 network interface and 30 IP addresses. If 5 pods are placed on the node and 5 free IP addresses are removed from the IP address warm pool, then ipamD attempts to allocate more interfaces until WARM_ENI_TARGET free interfaces are available on the node. If WARM_IP_TARGET is set, then this environment variable is ignored and the WARM_IP_TARGET behavior is used instead.


Type: Integer

Default: None

Specifies the number of free IP addresses that the ipamD daemon should attempt to keep available for pod assignment on the node. For example, if WARM_IP_TARGET is set to 10, then ipamD attempts to keep 10 free IP addresses available at all times. If the elastic network interfaces on the node are unable to provide these free addresses, ipamD attempts to allocate more interfaces until WARM_IP_TARGET free IP addresses are available. If both WARM_IP_TARGET and MINIMUM_IP_TARGET are set, ipamD will attempt to meet both constraints. This environment variable overrides WARM_ENI_TARGET behavior.


Type: Integer

Default: None

Specifies the number of total IP addresses that the ipamD daemon should attempt to allocate for pod assignment on the node. MINIMUM_IP_TARGET behaves identically to WARM_IP_TARGET except that instead of setting a target number of free IP addresses to keep available at all times, it sets a target number for a floor on how many total IP addresses are allocated.

MINIMUM_IP_TARGET is for pre-scaling, WARM_IP_TARGET is for dynamic scaling. For example, suppose a cluster has an expected pod density of approximately 30 pods per node. If WARM_IP_TARGET is set to 30 to ensure there are enough IPs allocated up front by the CNI, then 30 pods are deployed to the node, the CNI will allocate an additional 30 IPs, for a total of 60, accelerating IP exhaustion in the relevant subnets. If instead MINIMUM_IP_TARGET is set to 30 and WARM_IP_TARGET to 2, after the 30 pods are deployed the CNI would allocate an additional 2 IPs. This still provides elasticity, but uses roughly half as many IPs as using WARM_IP_TARGET alone (32 IPs vs 60 IPs).

This also improves reliability of the EKS cluster by reducing the number of calls necessary to allocate or deallocate private IPs, which may be throttled, especially at scaling-related times.


Type: Integer

Default: None

Specifies the maximum number of ENIs that will be attached to the node. When MAX_ENI is unset or 0 (or lower), the setting is not used, and the maximum number of ENIs is always equal to the maximum number for the instance type in question. Even when MAX_ENI is a positive number, it is limited by the maximum number for the instance type.


Type: String

Default: DEBUG

Valid Values: trace, debug, info, warn, error, critical or off. (Not case sensitive)

Specifies the loglevel for ipamd.


Type: String

Default: Unset

Valid Values: stdout or a file path

Specifies where to write the logging output. Either to stdout or to override the default file.


Type: String


Specifies the bind address for the introspection endpoint.

A Unix Domain Socket can be specified with the unix: prefix before the socket path.


Type: Boolean

Default: false

Specifies whether introspection endpoints are disabled on a worker node. Setting this to true will reduce the debugging information we can get from the node when running the script.


Type: Boolean

Default: false

Specifies whether the prometheus metrics endpoint is disabled or not for ipamd. By default metrics are published on :61678/metrics.


Type: String

Default: eni

Specifies the veth prefix used to generate the host-side veth device name for the CNI. The prefix can be at most 4 characters long.


Type: String

Default: {}

Example values: {"tag_key": "tag_val"}

Metadata applied to ENI help you categorize and organize your resources for billing or other purposes. Each tag consists of a custom-defined key and an optional value. Tag keys can have a maximum character length of 128 characters. Tag values can have a maximum length of 256 characters. These tags will be added to all ENIs on the host.

Important: Custom tags should not contain prefix as it is reserved. If the tag has string, tag addition will ignored.


L-IPAMD(aws-node daemonSet) running on every worker node requires access to kubernetes API server. If it can not reach kubernetes API server, ipamD will exit and CNI will not be able to get any IP address for Pods. Here is a way to confirm if L-IPAMD has access to the kubernetes API server.

# find out kubernetes service IP, e.g.
kubectl get svc kubernetes
kubernetes   ClusterIP   <none>        443/TCP   29d

# ssh into worker node, check if worker node can reach API server
telnet 443
Connected to
Escape character is '^]'.  <-------- kubernetes API server is reachable

Security disclosures

If you think you’ve found a potential security issue, please do not post it in the Issues. Instead, please follow the instructions here or email AWS security directly.



You can’t perform that action at this time.