The switching capabilities of bridges, routers, and hubs add new power to LANs and make it easier than ever to connect them to enterprise WANs
John Bryan
As more people participate in client/server environments, the computing model shifts from a physical workgroup to a project orientation. In the project-oriented environment, there is a need for networks that satisfy the end user's demands for bandwidth, access, and security, while still providing management capability at the corporate level.
Currently, networks are generally expanded using router, bridge, and fast-backbone technologies. All enable more diverse network topologies and improve performance through segmentation and increased bandwidth. Routers and bridges have always used switching technology; only recently has the switching
hub brought switching to the workgroup level.
Until now, LANs with access to shared media were standard. A 10-Mbps Ethernet network and 4- and 16-Mbps token-ring networks serve most business applications well, but they have inherent limitations. As more nodes are added, the available individual bandwidth decreases.
For example, in a 10-node 10Base-T network, each individual node can theoretically send or receive 10 Mbps but in practice have only 1 Mbps available. With little or no network traffic, there's no problem. But delays increase as nodes become more active or the network grows, as each node contends for a piece of the total bandwidth. With this example, if all 10 nodes are transmitting, total available bandwidth may be as low as 4 Mbps. The same is true for token-ring networks; the more nodes contending for the token, the less time any individual node has to transmit.
Of course, most LANs have more than 10 nodes. Burgeoning traffic volumes and graphical applications mean that chan
ges must be made. At least two approaches have already been tried. The first was to increase network speed. FDDI (Fiber Distributed Data Interface) runs at 100 Mbps but suffers from other problems, including cost and complexity (see ``All-Terrain Networking,'' August 1993 BYTE). But shared access, even at 100 Mbps, will eventually lead to the same problems.
Another solution is to divide the network into separate segments connected by bridges or routers. Each segment is shared, but communications can occur simultaneously in different segments, easing congestion. This is the right approach, but network growth will eventually cause the same traffic problems. The logical solution is to reduce each segment to one node, and that's where switching comes in.
Switching in LANs is exciting because it reverses the bandwidth equation. In the earlier 10Base-T network example, 10 nodes got 1 Mbps each. However, with a switched hub, each node can realize its full 10-Mbps potential; the total segment capacity w
ould be close to 100 Mbps. Except for some overhead, this relationship is maintained right up to the hub's bandwidth capacity.
As a result, segment traffic in legacy networks is greatly augmented without having to change NICs (network interface cards) or cabling. This has been the primary focus of switched hubs for LANs. In WANs (wide-area networks), the focus is more on reliability and efficiency.
Today, switching technology is being used for all types of network protocols at the LAN and WAN levels. Vendors offer switching products for Ethernet at 10 and 100 Mbps, for token ring at 4 and 16 Mbps, and for ATM (Asynchronous Transfer Mode). In addition, Fibre Channel is a switched technology being developed for high-end workstation use (see ``Fibre Channel Speeds Up,'' August BYTE).
Bridges and Routers--How Switching Works
Switching is required whenever a network signal must move between carriers (e.g., Ethernet to X.25) or change speeds (e.g., 10-Mbps Ethernet to 100-Mbps FDDI). Two
primary categories of switching are packet and circuit. Packet switching uses a fixed-length data cell (e.g., an entire Ethernet frame, packets of arbitrary size, or ATM ``cells''). In circuit switching, the carrying element is switched, generally for the duration of an entire message.
Bridges use MAC (media access control) address information from the data- link layer of the OSI (Open Systems Interconnection) network model to determine an incoming packet's destination. When a bridge encounters an unknown MAC address, it sends that packet to all its ports; this is called flooding, and it can present both traffic and security problems. Bridges learn address/port relationships dynamically--if a new address is real, it gets added to the correlation table. But if a bridge hasn't yet assimilated a new address, packets may be sent to users who aren't supposed to receive them.
The MAC layer also reserves certain addresses as ``broadcast'' designators--packets that are supposed to go to everyone on the
network. Unfortunately, this interrupts all other messaging. Used too frequently, the result is a broadcast storm that can thoroughly disrupt a network.
Routers take a different approach, using the OSI network layer, which establishes a different addressing scheme for each type of network protocol. To be compatible with multiple standards, routers have more complex switching logic than bridges, with associated delays and added cost.
Neither the classical bridge nor the router is a connection-oriented switch, and this presents problems. Connectionless switches do not maintain an end-to-end link, thus eliminating the possibility of data streams and isochronous traffic.
What's in a Name?
To allow each of its ports to operate at its maximum rate, a switch must handle the total potential traffic volume of all attached nodes, as well as switching overhead, the extra code bits assigned to each data-traffic unit. These traffic units might be variable-length Ethernet frames or constant-lengt
h packets, such as 53-byte ATM cells. Xedia's Smartswitch architecture, for instance, uses a 64-byte packet for Ethernet switching. For variable-length packets, software must determine the beginning and end; this introduces complexity and takes longer. Fixed-length packet switching can be implemented completely in hardware, which is faster.
Switch type can be defined by function, by the traffic units switched, or by hardware configuration. Most people are familiar with functional labels--bridge, router, and switched hub. The traffic-type label--Frame Relay, Ethernet, token ring, Fibre Channel, and ATM cell--tells you the intended use, but it gives no clue as to what's happening inside the box. What's happening inside the box determines the relative advantages and disadvantages, as well as the adaptability to future needs.
Several vendors produce so-called super hubs (or, as SynOptics Communications calls its Model 5000, Network Center Hubs)--switches that support some combination of Ethernet, to
ken-ring, FDDI, ATM, and WAN connections. Cabletron Systems, 3Com, and others have products in this category, but only Cabletron currently supports all those topologies.
Approaches to Packet Switching
Generally speaking, there are three types of packet-switching hardware: shared-memory systems, shared-bus designs, and multistage matrix switches. But simple categories don't define all the products, and vendors' definitions don't adhere to any particular standard. No one design seems dominant for any given type of switch. As a result, a savvy network manager needs to assimilate an intimidating array of information before making decisions. In fact, price-per-port costs, even within a single category, range all over the place.
Shared-memory and shared-bus architectures have common elements. Both buffer I/O in memory that's connected to the switch logic by a bus. The shared-memory design generally relies on a logic-managed common pool of memory for switching between ports, while the shared-bus
design uses a bus whose bandwidth is significantly higher than the accumulated demands of all attached ports.
A multistage matrix switch is an array of switching nodes, each with two inputs, two outputs, and a control line. Any input can be directed to any output. A number of ATM switches use the multistage-matrix-switch design. Because ATM cells are fixed-length, each node switch in the matrix can be a simple, fast, low-cost part such as an ASIC (application-specific IC) or FPGA (field programmable gate array). Xilinx (San Jose, CA), the leading FPGA vendor, has done extensive research on multistage matrix switches, such as a Banyan or Benes, implemented in a single FPGA.
ATM switches are not always based on the multistage-matrix-switch design. Fore Systems (Warrendale, PA) uses a variation on the shared-bus theme in its contentionless time-division architecture, and it has sold more ATM switches than anyone else--though that may also be a function of availability. Fore cites four advantages fo
r TDM (time-division multiplexing): It is nonblocking, has a predictably low latency, provides integral multicasting without copying, and will support flexible interface speeds.
Nonblocking means it won't prevent an incoming cell from entering the switch fabric, because the fabric has greater capacity than any possible load. Multistage-matrix-switch vendors have reduced, but not eliminated, traffic contention in the switch fabric. Fore's ForeRunner ASX-100 offers the lowest latency times for any ATM switch, though matrix switches are almost as fast. To perform a multicast, a multistage matrix switch has to copy cells at appropriate switch nodes until it sends packets to the correct output ports; this adds traffic to the switching fabric. The ForeRunner sends only one copy of any cell onto its bus, and ports take it off as necessary.
WANs at Speed
Switching is an expensive proposition, though cheaper than replacing NICs and rewiring a network to gain speed. But for wide-area applications, l
imiting costs are not for hardware but for data transmission. For instance, maintaining a connection between Chicago and Minneapolis on a T32 link--a 32-Mbps link equivalent to T1 lines--costs close to $30,000 per month.
The trick is to provide bandwidth on demand at all levels. Currently, most WAN connections are provided through PVCs (permanent virtual connections), whose parameters--but not necessarily routes--are established in advance. On the other hand, SVCs (switched virtual circuits) provide resources as required, with billing according to use. SMDS (Switched MultiMegabit Data Service) and Frame Relay vendors are expanding their SVC offerings, but some time-sensitive applications, like voice and full-motion video, really work best with the small cell size and short latency of ATM. Most WAN connections are supplied by routers, though the trend is moving from the circuit-switched router to the packet-switching variety.
Virtual Networking
One of the interesting and exciting possibilit
ies of switch technology is virtual networking--using software to create network user segments regardless of physical location or connection type. The goal is to make participants appear local to each other to facilitate communication.
For a physical network, the network administrator can map out workgroups, establish connections, and collect management data. In a virtual network, this is not so simple. In fact, at this point, virtual networking seems to introduce more problems for both network managers and users than it solves. For instance, how do you maintain routing address conventions if a user belongs to two or more networks? For that matter, can a single node exist in two virtual networks simultaneously? How do you guarantee data security?
Where the Technology's Heading
Switched networks will take us into the next generation of communications. Integrated voice, data, and imaging over any distance is only possible with the performance that dedicated bandwidth can provide.
The
promises of the Information Highway will come true only when people can economically deliver this power to every desktop. Switching is one of the key components that will make that happen.
John Bryan is a freelance technology writer and consultant based in San Jose, California. You can reach him on BIX c/o ``editors.''
LANS Make The Switch
