Several proposed technologies promise better performance for tomorrow's LANs. The promises create difficult choices for LAN owners who need better performance today, but don't want to buy equipment that will be non-standard or obsolete when tomorrow's technologies are standardized.
One performance-upgrade technology, Ethernet switching, exploits today's standards to improve the performance of LAN equipment already installed. With Ethernet switching, LAN owners can improve performance today, while making minimal equipment changes and keeping options open for tomorrow.
Several other performance-upgrade technologies compete for the LAN owners' attention: FDDI (Fiber Distributed Data Interface), FDDI over UTP (unshielded twisted pair wiring), Fast Ethernet, and ATM (Asynchronous Transfer Mode).
FDDI and its unshielded twisted pair version, FDDI over UTP, conform to today's standards. Many FDDI and FDDI over UTP products are available. They are appropriate for file server and backbone connections, but provide more capability than most desktops require.
Full duplex Ethernet, Fast Ethernet (100Base-TX and 100VG-AnyLAN) and ATM are the subject of standardization activities. Pre-standard products from different vendors may not interoperate. Even so, pre-standard LAN products may provide needed benefits in particular circumstances, especially in the backbone, by providing better performance, lower cost, simpler management, or some other advantage over mature standards.
LAN traffic increases as more computers are connected to existing LANs. If this were the only source of increased traffic, LAN upgrades could be confined to backbones that interconnect multiple LANs. Current bridges and routers could divide large LANs into multiple LANs at no great increase in cost per desktop, keeping the ratio of computers to individual LANs at a level that would produce acceptable performance. New applications, however, create a more significant kind of traffic increase: they cause the traffic per desktop to increase.
The shift to client-server computing splits applications between server computers and client (desktop) computers. The resulting communication over the LAN between clients and a server produces much greater LAN traffic per desktop than applications centered mainly in the desktop or in the server.
To reduce software management costs and administer copyright obligations, many companies now install generally used software, like word processors, spreadsheets, and Windows, only on servers. Each time a desktop computer invokes a program stored on a server, the LAN transfers the program to the desktop computer, producing greater per-desktop LAN traffic than if the program were stored at the desktop.
The shift to graphical user interfaces, and to greater use of their capabilities, means that documents stored on file servers include more illustrations or pictures (sometimes moving pictures) and sound bites. The movement of these larger documents between file servers and desktops creates increased per-desktop traffic.
Increasing traffic means that many LANs must be upgraded if unacceptable delays are to be avoided.
LANs need greater carrying capacity (bandwidth) for each desktop computer. Current LANs share a total bandwidth of 10 megabits per second (Mbps) among many desktops and one or more file servers. Simply installing a faster shared LAN medium may fail to solve a performance problem. A fast shared medium capable of supporting several hundred users would require a speed of many hundred Mbps. LAN media with such speeds would be extremely expensive if connected directly to each desktop, and are not available anyway. LAN media operating at 100 Mbps are available, but 100 Mbps is not enough total capacity for larger networks. In small networks, where 100 Mbps total capacity would be sufficient, a 100 Mbps LAN medium connected to each desktop provides more than the required capacity and a higher level of expense.
Reducing the degree of sharing of the 10 Mbps bandwidth by dedicating a LAN to each small group of desktops, or to each individual desktop in some cases, is sufficient for the next few years. An individual file server, however, may generate 30 Mbps of traffic or more. A cost-effective but high-performance LAN must provide higher speed connections (100 Mbps) for file servers, and provide lower speed connections (10 Mbps) for desktops.
A LAN switch is a device that reduces the sharing of LAN media by keeping traffic to the segments of the media for which it is destined, rather than wasting bandwidth on other segments. A LAN switch divides a LAN into many segments, thereby dividing a large volume of traffic into many smaller traffic flows. In a LAN switch, the shared medium is a high-speed backplane or other switching "fabric." Since the switching fabric is located entirely within a single cabinet, the switch provides high bandwidth for least expense. The LAN segments that connect the high-capacity switch to desktop computers have relatively low capacity and cost, because there are many segments to carry a large total traffic volume. The lower capacity of the LAN medium, since it is shared among fewer computers, is more than adequate for small work groups or individual desktops. This division of the LAN into small segments is generally called microsegmentation.
There are two broad categories of switching products being offered to LAN owners: Ethernet switching and ATM. Ethernet switching improves Ethernet performance without introducing a need for new standards. ATM is a set of proposed standards for data transport and switching. ATM promises to integrate LANs, WANs, voice, and video. Ethernet switching and ATM do not exclude each other Ethernet switching will be used as a low-cost interconnection between desktop computers and ATM backbones. Both Ethernet and ATM can be switched entirely by procedures implemented in hardware; software is not required for data movement.
Ethernet (standardized as IEEE 802.3 and ISO 8803) was the first LAN technology, and has always been the most widely used. There are more than 30 million Ethernet computer connections in use. Ethernet networks operate at a data transmission rate of 10 Mbps.
Early Ethernet standards were based on coaxial cable. The coaxial cable ran directly from computer to computer, rather than radially from a central point to each computer. This arrangement created many problems, since a fault at any connection point affected the whole network. Standardization of the 10Base-T version of Ethernet eliminated many problems that were due to cabling.
The 10Base-T Ethernet standard specifies centralized repeaters, or hubs, and unshielded twisted pair wiring radiating from the hubs to each computer. UTP is similar to building telephone wire, and has great practical and economic advantages over coaxial cable. The 10Base-T standard greatly improved the fault detection and fault isolation properties of Ethernet. For example, a standard 10Base-T repeater hub detects incorrect behavior of a station and isolates it, preventing it from disrupting the network.
A single Ethernet LAN is generally called a segment. Large LANs may be composed of multiple segments. Each segment may have many computers attached to it. Ethernet is a broadcast, or shared, medium; data transmissions from one computer propagate everywhere on the segment. The receiver in the network adapter in each computer selects frames of data meant for it by inspecting the addresses in all frames on the segment. With a shared medium, the cable between an individual computer and its connection point in the equipment closet carries traffic for all computers on the segment.
Switching offers a way of using the Ethernet standard that greatly increases its performance without requiring changes to cabling, network adapters, or computer software. An Ethernet switch divides the LAN into many small segments. Instead of sharing the limited capacity (10 Mbps) of Ethernet with many computers, each computer, or each small group of computers, can have a dedicated 10 Mbps segment connected to a high-capacity Ethernet switch (40 to 1,000 Mbps.)
An Ethernet switch is a device with multiple Ethernet connections, or ports. An Ethernet switch also needs some higher speed ports for file server or backbone connections via FDDI, Fast Ethernet, or ATM. The function of an Ethernet switch is similar to the function of a bridge: it transfers frames of data between LAN segments, confining traffic between destinations on the same LAN segment to that LAN segment, and confining traffic between destinations on two LAN segments to those two LAN segments. The functions performed by a switch may be identical to those of a bridge, although some switches perform only selected pieces of the bridging function. The term "switch" is usually reserved for a device that has many Ethernet ports, all or most of which can accept or transmit frames simultaneously at the full rate of the connected LAN media (so-called "wire rate" ports).
An Ethernet switch interconnects a large number of ports at low cost by moving frames of data between ports entirely by electronic logic. Microprocessors and software do not participate in basic data movement. Ethernet switching procedures can be encapsulated entirely within Application-Specific Integrated Circuits (ASICs). Bridges and routers, in contrast, typically use high performance RISC microprocessors to move data frames. Frame movement by microprocessor is either more expensive or slower, and requires larger devices, larger cabinets for a given number of Ethernet segments, and more electric power.
There are two broad classes of Ethernet switching procedures:
A cut-through switch starts to forward a frame to its destination before receiving the end of the frame. A cut-through switch can achieve the lowest forwarding delays, but it propagates errors from one LAN to another, because errors can only be detected at the end of each frame. Error propagation is a particular concern with Ethernet, since correct operation of the Ethernet protocol generates corrupted and truncated frames (as collisions occur during contention for the right to transmit).
A store-and-forward switch receives a whole frame before starting to forward it to its destination. By waiting for the end of the frame, a store-and-forward switch can verify that a correct frame is available for transmission before disturbing a destination LAN by transmitting a corrupted or truncated frame.
A store-and-forward procedure is required whenever a frame must be moved from a low-speed LAN to a higher speed LAN. FDDI and Fast Ethernet do not permit gaps in frames. To ensure that there are no gaps, a whole frame must be received from a lower speed LAN before beginning to transmit it on a higher speed LAN. A cut-through switch must change to a store-and-forward procedure when moving frames between LANs of different speeds. Because most frames move between LANs of different speeds, the necessity to change procedures eliminates the delay advantage of the cut-through procedure. Most LAN owners aim for this kind of environment, to achieve the best file server performance with 100 Mbps connections while achieving the best economy with 10 Mbps Ethernet connections.
Even where cut-through can be used, packet-burst (Novell NCP), or windowed (TCP), protocols greatly reduce the significance of slightly lower switching delay, giving store-and-forward essentially the same level of throughput. Few users, when they request a file from a server, care if reception of the beginning of the file is delayed by a few hundred microseconds; most will object if reception of the end of the file is delayed by seconds. Packet burst, or windowed, protocols prevent an extra delay of a few hundred microseconds in the propagation time of individual packets from being transformed into an increase of seconds in the duration of a file transfer. A packet-burst protocol transmits several or many packets before requiring acknowledgment of previously sent packets. Older protocols required an acknowledgment for every packet, which amplified the effect of delays in the path between the client and the file server.
A store-and-forward switch can achieve the lowest delay theoretically possible for a store-and-forward procedure by determining how to route each frame during reception of the frame. Overlapping of reception and address processing means that the switch can be ready to start forwarding immediately after reception of the end of the frame.
Three proposals for speeding up Ethernet are working their way through the IEEE standards process: full duplex Ethernet, and two 100 Mbps "Fast Ethernets" which are referred to individually as 100Base-TX and 100VG-AnyLAN.
Full duplex Ethernet provides a capacity increase rather than a speed increase, and operates only as a direct connection, or link, between two computers, or between a computer and a switch. On a full duplex link, traffic can flow in both directions simultaneously. Standard Ethernet is a half duplex shared medium, on which traffic can flow in only one direction at a time. If the offered traffic load most often flows in one direction at any one time, the full duplex link offers little advantage over a half duplex link of the same speed. Full duplex Ethernet offers a substantial performance improvement for links between two switches, where the offered traffic flow is usually symmetrical. A full duplex link may offer improvement for a connection between a file server and a switch, where most traffic flows in one direction (from the server through the switch to the clients), but substantial traffic still flows in the other direction (from clients through the switch to the server). For a connection between a workstation and a switch on which almost all traffic consists of solicited transfers from a file server, full duplex Ethernet offers little improvement.
The two Fast Ethernets, 100Base-TX and 100VG-AnyLAN, aim to get 100 Mbps to the desktop over unshielded twisted pair cables, at very low cost relative to FDDI or FDDI over UTP. Both Fast Ethernets use UTP cables, although both are standardizing fiber-optic cables as well.
Back to the KMJ White Paper Page
Back to the KMJ Home Page