The core job of an ISP is essentially to forward Internet traffic. Physically, this is done with devices called "routers", who work at the "network" layer (also called "layer 3") of the OSI and TCP/IP models. An IXP is a physical infrastructure where several ISP networks can be interconnected. This means that there has to be a connection between all participating ISP routers. This has been done in the past by physically connecting (i.e., at the so-called "layer 1") the routers (for example, with a 10BASE2 cable); however, most IXPs currently connect routers at the data link layer ("layer 2") by cabling them to Ethernet devices and letting them share a Local Area Network, called "Peering LAN".
The colocation
It is clear that, to run an Internet Exchange, physical space is needed in order to house the core infrastructure of the exchange. The IXP has to find space in a colocation providing at least power and cooling to the IXP devices. Possibly the main value of the IXP is resilience. It is usually valuable for an IXP to find a colocation that can offer improved services, such as redundant power feeds and the presence of an Uninterruptible Power Supply (UPS) and a power generator. This greatly reduces risks of downtime for the IXP.
Other colocation assets that are usually much appreciated by IXP operators are: restricted access to the premises (to minimize risks of strangers tampering with the IXP equipment), fire prevention and extinguishing equipment, flood prevention measures. It is generally considered that a good choice for the colocation is in a diversely connected area: that is, an area with the presence of several ISPs (possibly needing interconnection to each other, and thus the presence of an IXP) and several carriers that could provide the connection from the colocation to the ISP premises. Often, the colocation is carrier-neutral (but it does not have to be).
The devices
Nearly all IXP currently operating use Ethernet devices to provide a flat Layer 2 network to interconnect ISP routers. A single device is usually sufficient; however, with this configuration, one failure would bring the entire IXP down. Depending on the size and number of sites the equipment used can be a plain Layer 2 single VLAN technology up to a routed backbone with virtual LAN services on top. The latter one provides much more flexibility, fail over functionality and traffic engineering possibilities, but the flip side is they are much complex to operate and debug.
It is highly recommended for the devices to have redundant power supplies. If there are two or more power feeds, each supply should be connected to a separate feed. In this way, a failure to one of the power feeds or one of the power supplies would not affect the operation of the IXP. A complete hardware failure to the device would still bring the entire Exchange to a halt. To avoid this as much as possible, IXP operators often decide to rely on modular devices, where it is possible to have redundant CPU and forwarding cards.
In some cases, IXPs opt to have two completely separate peering devices, with a separate peering LAN. Every ISP who wants the redundancy should then have two routers or at least two interfaces to be connected to the LANs. This "redundant" setup is a bit like having two IXPs in a single location; if one of them becomes unavailable for whatever reason, the other can take over. Some IXP operators also choose to use different vendors for the IXP devices, so that a flaw in the software may not affect both peering devices; however, this is a choice that depends on other factors as well, such as the choice between single-chassis or modular infrastructure. All these choices should be made by keeping in mind that the most critical asset of an IXP is resilience; and obviously taking into account the IXP environment, the needs and the available funding.
When an IXP grows, a single-device or even double-device configuration may become insufficient; it may then become useful or necessary to add more devices to the infrastructure. These devices can be arranged in different topologies, the most common being the "star" topology (i.e. one or more "core" devices to whom several smaller "access" devices are connected) and the "ring" topology (where every device is connected to two "neighbor" devices, thereby forming a ring).
To improve redundancy and diversity of access to its infrastructure, the IXP may choose to install peering devices in more than one physical location. This should be planned so that "disaster" in one of the locations would not affect the operations in other locations.
