Gigabit Ethernet network transceivers augment Fibre Channel platform
Use of the existing Fibre Channel physical interface running at 1.25 Gbits/sec and its robust 8B/10B encoding scheme can markedly reduce Gigabit Ethernet transceiver development cost and time-to-market
Patrick van eijk
vitesse semiconductor Corp.
The increasing local area network (LAN) loads from faster computers, laser printers, scanners, video servers, high-speed graphical displays, and other high-resolution imaging applications have spurred the ieee 802.3 Gigabit Ethernet Interim Committee into action to develop a Gigabit Ethernet data transmission standard for speeds beyond the 100-Mbit/sec rates for Fast Ethernet 100Base-T and 100Base-VG AnyLAN.
To minimize risk and improve product time-to-market, the committee plans to leverage the existing Fibre Channel physical (PHY) transceiver designs, which run at 1.25 Gbits/sec. This rate includes a 20% data coding overhead. If the committee adapts the existing media access control (MAC) protocol, management information base (MIB), and the frame format and logical link control (LLC) interface, the Gigabit Ethernet standard should maintain maximum compatibility with the installed base of more than 60 million LAN nodes.
Fibre Channel, a point-to-point, serial protocol specified by the ansi X3T11 Committee, has been used successfully for several years. It is a layered protocol, loosely based on the OSI Network System Reference Model. The lowest layer--the PHY--dictates several different serial transmission rates, of which the most popular is the 1.0625-Gbit/sec, or full-speed, rate.
The Fibre Channel industry specifies 10- and 20-bit transistor-transistor-logic (TTL) parallel interfaces, which are serialized to the 1.0625-Gbit/sec data stream. At the receiving end, the serial 1.0625-Gbit/sec data is deserialized back to 10 or 20 bits. The highest integrated physical-layer device available is a 10-bit, 1.0625-Gbit/sec transceiver that has been developed for the dual-channel disk drive market. Because disk drives represent the gating items in the resurgent Fibre Channel market, transceiver volumes are expected to soon push prices to less than $15 per channel.
In fact, the proven transceiver functionality, performance and cost are what attracted the ieee committee to consider Fibre Channel as its method of implementing the Gigabit Ethernet physical layer. Furthermore, multiple vendors have committed to providing functionally equivalent transceivers at 1.25 Gbits/sec for the Gigabit Ethernet market. The adapted gigabit transceiver technology enables the implementation of three different Gigabit Ethernet links: 1000Base-T, full-duplex, and 1000Base-VG.
As current 10Base-T shared hub-and-spoke networks migrate to switched 10Base-T and 100Base-T networks to avoid the performance degradation experienced by users of shared-media LANs, the proposed Gigabit Network link methods, such as switch-to-switch and server-to-switch, are anticipated to prevent future network bottlenecks. Indeed, many networks have already migrated to such a collapsed backbone architecture (see Fig. 1). In this approach, the 100-Mbit/sec shared Fiber Distributed Data Interface backbone is replaced by a switch containing both 10- and 100-Mbit/sec ports, thereby permitting seamless upward migration.
Because gigabit switching is being deployed from the top down, starting at the server and router connections, it is expected to carve a path toward the end-user`s desktop. Meanwhile, as the price of switched networks decreases, more and more 10Base-T and 100Base-T repeaters are being replaced with Fast Ethernet switches interconnected by gigabit links (see Fig. 2). A fully employed switched network architecture, with a multigigabit aggregate bandwidth, is anticipated to support full-motion multimedia, videoconferencing, medical imaging, simulation and modeling, and other high-speed imaging applications.
Bidirectional operation
In switched-architecture networks of the future, most links are expected to be gigabit links in full-duplex operation. With data flowing simultaneously in both directions across the link, a 100% improvement in throughput (compared to a simplex link) should be delivered, assuming that traffic is evenly balanced between the send and receive ends.
However, using a carrier-sense multiple access/collision detect (csma/cd) MAC protocol for 1000Base-T links in conjunction with the current Ethernet minimum packet size (512 bits) would result in an unacceptably small domain diameter; that is, the link distance would be short. For example, in an Ethernet collision domain, the round-trip propagation delay must not exceed 512 bit times, which amounts to 409.6 nsec in a 1000Base-T collision domain. Assuming a 5.0-nsec cable delay (for fiber) and a combined repeater and data terminal device delay of 100 bit times, such as incurred by a personal computer, workstation, or server, the best-case collision-domain diameter is 32.96 m.
This distance value means that two data terminal devices talking to each other over a shared 1000Base-T link through a repeater can have a combined cable length of just 32.96 m, or an individual cable length of just over 16 m. With the average length of most horizontal wiring (from a repeater in a telecommunications closet to data- terminating equipment in a work area) far exceeding 16 m, a 1000Base-T shared network architecture based on existing Ethernet packet formats and sizes is unlikely.
In comparison, the maximum link limit in a shared 100Base-T network is 100 m for 100Base-T4 (unshielded twisted-pair [UTP] copper, Category 3, 4 or 5), 100 m for 100Base-TX (UTP Category 5 or shielded twisted-pair [STP]), and 412 m for 100Base-FX (multimode 62.5/125-micron fiber-optic cable only). Therefore, to run a csma/cd MAC protocol on a 1000Base-T link, network planners must either keep the individual link distance below 16 m or change the minimum Ethernet packet size by adding dummy packet fillers that would be stripped off at the receiving end. This method would, of course, have an adverse effect on the efficiency of a network where the traffic consists of a high content of minimum size packets. The ieee 802.3 Committee`s High-speed Study Group (hssg) is evaluating several proposals on how to increase the minimum link distance.
Fast Ethernet (100Base-T) has narrowed the gap between the anticipated high-speed switching and transport of Asynchronous Transfer Mode (ATM) technology and today`s available network performance. Consequently, widespread deployment of 155- or 622-Mbit/sec ATM technology to the desktop has been pushed out to the year 2000 or beyond, according to industry analysts. ATM technology at the LAN level is an attractive feature because of its perceived capability to provide wide area network (WAN) ATM connectivity after fully ATM-based WANs are deployed. For now, network operators have to choose between ATM, with its perceived future WAN capability, and switched-architecture 10Base-T and 100Base-T networks with 100Base-T and 1000Base-T switch-to-switch and switch-to-server links. The latter links come with greater availability, range of options and configurations, and software support.
Nearly all industry analysts agree that ATM technology will eventually prove to be the connectivity solution of choice for WANs, delivering the lowest-cost long-distance transmission services. Just like Ethernet, ATM supports all popular media, such as UTP Category 3, 4 and 5 copper, 150-ohm STP copper, and singlemode and multimode fiber. A switched architecture combined with high-speed links should provide nearly unlimited scalability and flexibility.
Although ATM cells provide a short time delay, which is advantageous not only in the WAN arena, but also in LAN applications, the ATM standards remain incomplete. Mechanisms for monitoring and policing traffic to prevent internal switch buffer overflow in public networks are especially lacking.
Like ATM, Ethernet also offers switching. A 100Base-T switched architecture network with gigabit-per-second switch-to-switch and switch-to-server links should offer a solid foundation for multimedia transport. Such a network, combined with emerging flow control mechanisms for on-time delivery of multimedia traffic--such as real-time transfer protocol (RTP), real-time transfer control protocol (rtcp), and resource reservation protocol (rsvp)--which smooth out the inherently jerky Ethernet traffic patterns, appears to be a serious ATM competitor in the LAN arena.
For a Gigabit Ethernet link, serial data propagating at 1.25 Gbits/sec cannot be transported over UTP. Instead, a 150-ohm coaxial cable or a fiber-optic cable must be used. Fiber is more attractive because it is immune to electromagnetic interference, and fiber cable lengths can be longer than copper cable. Since, initially, most Gigabit Ethernet links are expected to be installed from switch-to-switch or from switch-to-server, a link distance requirement exceeding 100 m appears necessary. In fact, the ieee 802.3 hssg has agreed upon and demonstrated a 500-m maximum link distance at 1.25 Gbits/sec over a 50-micron core multimode fiber, and a 200-m maximum link distance for a 62.5-micron core multimode fiber.
Using 150-ohm coaxial cable, existing Fibre Channel physical transceivers can only drive information for 30 m. However, this distance does not satisfy the existing link distance requirement for Fast Ethernet. However, when double-drive-strength pseudo-emitter-coupled logic (pecl) output drivers and an equalizer device are used, distances of 100 m or longer are achievable.
To date, network planners remain undecided as to whether the 8B/10B data encoding/decoding scheme (necessary to provide a high enough transition density to allow for the correct operation of the clock recovery circuit at the receiving end) should be used for both a fiber and a copper medium. To reduce the maximum frequency on a copper link, a different encoding scheme and non-binary signal levels could be used. However, this approach would result in the requirement for a new transceiver that would significantly delay the development of copper Gigabit Ethernet links.
Note that the only difference between the Fibre Channel physical transceiver, operating at 1.0625 Gbits/sec, and a potential Gigabit Ethernet transceiver is speed. If the ieee committee decides to implement the Gigabit Ethernet physical layer by using a Fibre Channel physical transceiver running at 1.25 Gbits/sec, then a complementary metal-oxide semiconductor (cmos) physical controller is required to implement 8B/10B data encoding/decoding (or a different encoding/decoding scheme), management registers, error checking and other logic that is needed to provide a modified media-independent interface (MII+), as shown in Fig. 3. The main difference between a Fast Ethernet and a Gigabit Ethernet MII+ is the number of parallel bits--eight in the case of 1000Base-T. Since 10-bit, 1.25-Gbit/sec transceivers are currently available, network planners can readily implement full-duplex Gigabit Ethernet adapter cards (see Fig. 4). u
Patrick van Eijk is a field applications engineer at Vitesse Semiconductor Corp., Sunnyvale, CA.