How to use a second Elastic Network Interface on the same subnet

Question

When I spin up an Amazon EC2 CentOS 7 server in, say, availability zone us-east-1a, the server is automatically assigned a primary private IP address on eth0, such as 172.31

Answer 1

The problem is asymmetric routing. A request to eth1 comes in eth1 and goes out eth0. The reply coming out on eth0 has a different IP address than in the request, and so it is dropped on the client side. The solution is to set up rules that allow responses to route through eth1.

First, make sure you have created an AMI of your server, because if you enter the wrong thing in following steps, you may lose all connectivity to the server and be unable to do anything but reboot it from the Amazon console web page.

Start off by setting the default route for each interface in separate tables:

ip route add default via 172.31.0.1 dev eth0 tab 1
ip route add default via 172.31.0.1 dev eth1 tab 2

To check those were properly added use:

ip route show table 1
ip route show table 2

Now you need to add rules that say to use the different tables depending on the source IP address:

ip rule add from 172.31.8.244/32 tab 1
ip rule add from 172.31.12.121/32 tab 2

You can check all of the rules with:

ip rule

You should now be able to connect to either IP address from a client machine. You can also use the bind option of SSH to connect from either interface on this server to a client machine:

ssh centos@client_ip_address -i mykey.pem  (uses the default, eth0)
ssh -b 172.31.12.121 centos@client_ip_address -i mykey.pem  (uses eth1)
ssh -b 172.31.8.244 centos@client_ip_address -i mykey.pem  (uses eth0)

You can use both interfaces to connect to other EC2 servers in the same availability zone and for any interface that has a Public IP assigned to it, you can connect to the outside world or to other EC2 servers in the same VPC, even if they are in different availability zones.

But what if you want to connect to other EC2 servers that are in the same VPC but different availability zones? In other words, servers in the same data center. The problem is that the Private IP address is masked at 20 bits, which confines you to one availability zone. So for datacenter us-east-1 you have:

us-east-1a:  172.31.0.0/20
us-east-1b:  172.31.16.0/20
us-east-1d:  172.31.48.0/20
us-east-1e:  172.31.32.0/20

To connect across availability zones in one VPC and in one datacenter you need a 16-bit mask. ip addr will show:

inet 172.31.12.121/20 brd 172.31.31.255 scope dynamic eth1

If losf -n | egrep 172.31.12.121 shows you that this address is not in use you can add the new mask and delete the old. Note that the broadcast address has to change at the same time the mask changes:

ip addr add 172.31.12.121/16 dev eth1 brd 172.31.255.255
ip addr del 172.31.12.121/20 dev eth1

Now you should be able to connect from an EC2 server in availability zone A to another host in availability zone B, so long as they are in the same VPC, even if they do not have Public IP addresses.

Troubleshooting:

If you are having problems, try resetting both interfaces, which will remove any manual twiddling you have done. First copy /etc/sysconfig/network-scripts/ifcfg-eth0 to /etc/sysconfig/network-scripts/ifcfg-eth1, editing the second file to change the DEVICE from eth0 to eth1. Then add a line to /etc/sysconfig/network which says GATEWAYDEV=eth0. Finally, run /etc/init.d/network restart (no, it should not disconnect you). Then start over with the above commands.