Smarter and Faster IP Connections

New IP-switch designs help move low-latency data such as sound and video through large networks.

Mick Seaman

Not so long ago, 80 percent of all network traffic was contained within common subnets. Today, the phenomenal growth of the Internet and business intranets has dramatically increased the amount of traffic that must be routed among separate subnets. Furthermore, network administrators who once had to worry far more about the reliability of data and little about when it arrived are now faced with demands for bidirectional audio and video. In these examples, it's expe cted that there will be a small, acceptable data loss, but the issue of when this real-time data arrives at the desktop via increasingly busy networks has become vital.

Unfortunately, current routing technologies are not suitable for cost-effective, multigigabit low-latency traffic. This means that most LANs use switching as the basis for high-speed traffic among subnets on a local network, but they use slower routers for moving data among subnets on different LANs. Thus, as data moves among subnets (an ability made possible by the routers), it can face unpredictable delays.

For these reasons, network managers want to design their LAN infrastructures on high-speed-switching architectures, because switches provide wire-speed forwarding between separate LAN segments while creating a single logical LAN between end systems. New solutions being brought to market by two leading network suppliers aim to provide the control-policy functions of routing with the wire-speed performance of switching.

IP Switching

Ipsilon's IP Switching establishes virtual circuits that bypass routers' Open Systems Interconnection (OSI) network level 3 layer using flow-matching techniques. In IP Switching, which is targeted at asynchronous transfer mode (ATM) networks, each IP node sets up a virtual channel on each of its ATM physical links to be used as the default forwarding channel. An ATM input port inside each switch receives incoming traffic on this default channel and sends it to Ipsilon's intelligent routing software in its switch controller. In addition to forwarding the packet over the default channel, the switch controller identifies the flow. A flow is a sequence of packets with the same point of origin, the same destination, the same protocol type, and other common characteristics.

The switch then performs a decision-making process to determine whether a flow should be routed or switched to a high-speed ATM virtual circuit. For a time-critical flow, the switch con troller establishes a virtual circuit, eliminating the need for further router processing, as shown in the figure "Ipsilon's IP Switching Mechanism."

While this architecture does result in performance improvements, there are several potential drawbacks to Ipsilon's switching solution. First, the architecture involves moving the router aside in favor of the Ipsilon switch controller, all in one step. Network managers may be unwilling to make such a change with the core piece of their networking infrastructure.

Second, there's IP Switching's flow orientation. While opening a virtual circuit makes sense in many cases, the technology relies on predictions from the switch controller whether to establish the circuit. For relatively small data transfers, opening the virtual circuit may not be worth the overhead that creating the virtual circuit imposes.

Finally, IP Switching is suited only for ATM network architectures. Few LAN backbones are solely ATM-based. Therefore, Ipsilo n's IP Switching technology is suitable for only a small segment of the marketplace.

Fast IP

The Fast IP protocol from 3Com (the author is an employee of 3Com) offers the performance of switching with the control of routing over all types of network backbone technologies, including Ethernet, Fast Ethernet, Gigabit Ethernet, Fiber Distributed Data Interface (FDDI), Token Ring, and ATM OC. Fast IP is applicable in both packet- and cell-based networks.

Fast IP is different from other IP-switching solutions in that it is initiated at the desktop, not in a router or switch. By equipping desktops and servers with the means to tell the network what they need and when they need it -- and then explicitly tagging the associated frames -- networks can implement the required quality of service policies without guessing or compromising performance by having to examine details in frames. Fast IP also reduces the number of layer 3 routing hops wherever possible, thus maintaining network simplicity and s peed, and reducing latency.

A Fast IP connection begins at the desktop system through a Next Hop Resolution Protocol (NHRP) request and response technique. NHRP uses source and destination media access control (MAC) addresses to establish a layer 2 connection. It also optionally uses tags defined under the IEEE-802.1q "Draft Standards for Virtual Bridged LANs," known as Group Address Registration Protocol (GARP).

The desktop addresses its first packet to the layer 3 router. The router forwards the packet to its destination, while applying common filter/firewall policies. When the server receives the packet, an NHRP response is sent via layer 2 directly to the originating desktop's address. If the response packet reaches its destination, it indicates that there is a directly switched path to the server. The desktop then uses the server's MAC address to communicate via layer 2, bypassing the routers, as shown in the figure "3Com's Fast IP Mechanism." If the response is not rece ived, the data flow continues to be routed as before.

In addition to simplifying management and enhancing speed by bypassing routers, Fast IP is based on several emerging standards, including IEEE-802.1q, Internet Engineering Task Force (IETF) NHRP, and IEEE-802.1p "Draft Standard for Traffic Class and Dynamic Multicast Filtering Services in Bridged LANs."

Fast IP is an affordable solution, being software-based. Because it is initiated and controlled solely by desktops and servers, it requires no changes to switches and routers. All that's needed to achieve Fast IP benefits is to add software to the appropriate systems. Client software and support for switches will be available from 3Com in the second half of this year. Fast IP client software will be bundled with certain PC network interface cards (NICs), and you can download it from 3Com's Web site ( http://www.3com.com ).

Migration Path

What's probably of interest to the network manager is that Fast IP offers a gradual migration path. It does not remove the router. It simply speeds up the router's performance. As mentioned earlier, it requires only software installation on the end systems (desktops and servers). No changes are necessary to the hardware or software of existing routers in the network to support Fast IP. Importantly, Fast IP interoperates with switches that don't support 802.1p, 802.1q, and NHRP. Thus, a manager can slowly upgrade the end systems without worrying about bringing down the network.

---

Ipsilon's IP Switching Mechanism

illustration_link (25 Kbytes)

You can send low-latency data directly through an ATM virtual channel, bypassing the IP controllers.

---

3Com's Fast IP Mechanism

illustration_link (27 Kbytes)

Fast IP uses standard network protocols to establish a layer 2 network link for low-latency data.

---

Mick Seaman is vice president and chief technical officer of 3Com's network systems operations. You can contact him by e-mail at editors@bix.com .