Does a router's ARP table contain the MAC address for all destination IPs?

Question

Let's say we have the following configuration.

WAN - Router -------------------- Switch --------- Client
                       |                     |
   Multiple switches connected to router     |
                                   Multiple Clients connected to each switch

When the router receives a packet, then the router would have to lookup its ARP table to find the appropriate MAC that corresponds to the IP address to create the frame to send off to the switch.

However, let's assume among the many switches there is a switch, let's call it S1, that is only connected to clients with IPs in range 123.123.123.0/24. This would mean that we can set up a routing rule at the router to route all traffic in range of 123.123.123.0/24 to be sent to the interface that leads to S1.

We can see that when routing, the router wouldn't need a routing table of every client. However, the ARP table would need a table of every client. That seems very inefficient to me. Is this really how it works? Wouldn't it make more sense for the Switch to be able to have its own ARP table so that you only need to look up a subset of clients?

To me, it would make more sense for the router here not to build a frame and send the packet that lacks L2 info to the switch, and then the switch builds a frame if the frame is incomplete. I know that switches operating in L2 does not want to look at the IP addresses, but I don't understand why it's designed that way.

Answer 1

It's not particularly inefficient, as long as the clients all fit in each switch's MAC table – the CAM "content-addressed memory" is pretty good at finding the MAC table entry efficiently.

Wouldn't it make more sense for the Switch to be able to have its own ARP table so that you only need to look up a subset of clients?

Then it would be a router, literally. What you suggest is often done, just with actual IP routing – the first router directs the packet to the MAC address of the second, knowing that the /24 is there.

Your example is inefficient because you designed it that way. A network isn't inherently limited to one router, and a router isn't limited to one internal network, so you don't have to have one huge subnet if you can avoid it – you can have the /24 as a completely separate subnet behind its own router if you want, or a VLAN managed by the same router but separated at L2 using 802.1Q tags.

(Where a "home" router has one LAN interface going into an integrated switch, enterprise routers have many independent interfaces to route between.)

On the other hand, when you really need efficiency, there's a third way to forward packets: label switching, such as MPLS. It allows an ingress router to pre-select the L2 path through a network, with the interior routers/switches only doing lookups in small label tables.