Elastic Fabric Adapter

For a list of instance types that support EFAs, see Supported instance types in the Amazon EC2 User Guide for Linux Instances.

$ aws ec2 describe-instance-types  --region us-east-1  --filters Name=network-info.efa-supported,Values=true  --query "InstanceTypes[*].[InstanceType]"  --output text | sort

For a list of operating systems that support EFA, see Supported operating systems.

The AMI has the EFA driver loaded.

The security group for the EFA must allows all inbound and outbound traffic to and from the security group itself.

All instances that use an EFA must be in the same cluster placement group.

The job definition must include a devices member with hostPath set to /dev/infiniband/uverbs0 to allow the EFA device to be passed through to the container. If containerPath is specified, it must also be set to /dev/infiniband/uverbs0. If permissions is set it must be set to READ | WRITE | MKNOD.

The location of the LinuxParameters members are different for multi-node parallel jobs and single-node container jobs. The following examples show the differences, but are missing required values.

{
  "jobDefinitionName": "EFA-MNP-JobDef",
  "type": "multinode",
  "nodeProperties": {
    ...
    "nodeRangeProperties": [
      {
        ...
        "container": {
          ...
          "linuxParameters": {
            "devices": [
              {
                "hostPath": "/dev/infiniband/uverbs0",
                "containerPath": "/dev/infiniband/uverbs0",
                "permissions": [
                    "READ", "WRITE", "MKNOD"
                ]
              },
            ],
          },
        },
      },
    ],
  },
}

{
  "jobDefinitionName": "EFA-Container-JobDef",
  "type": "container",
  ...
  "containerProperties": {
    ...
    "linuxParameters": {
      "devices": [
        {
          "hostPath": "/dev/infiniband/uverbs0",
        },
      ],
    },
  },
}