Ethernet Endurance Contest

Published in LAN Magazine   May 1996 (Now known as Network Magazine)

Buddy Shipley

 

   Although Ethernet originated as a coax-based shared media network it has evolved over the years in response to the ever-changing requirements of its users. To keep pace with the most recent cable plant demands the standards now accommodate additional media options including optical fiber and unshielded twisted pair. The motivating force driving the move to such media was the rapid and pervasive spread of LANs throughout businesses, governments, and other organizations worldwide, and the need to efficiently and economically manage and maintain them. The old, non-structured coax cable plants just could not provide this.

   Ethernet switching was developed to provide greater bandwidth availability to the stations and servers that need it. The emergence of cost effective LAN switches prompted the migration to low-density, and private or dedicated networking through micro-segmentation. As a significant new facet to the LAN industry, the switch market grew 390% last year -- from 413,000 ports shipped in 1994, to over 2 million ports shipped in 1995, according to IDC. Without such relatively inexpensive, low latency LAN switches, this step in Ethernet's evolution would not have been possible.

   The newest entrant into the Ethernet market, 802.3 100Base-T Fast Ethernet, has not yet made a significant impact on the LAN industry. Total high-speed LAN adapters shipped (including 100Base-T, 100VG AnyLAN, and ATM) approached one million for the year, according to IDC. The 100 Mbps products are starting to ramp up as more vendors jump into the game.

   Fast Ethernet provides an elegantly seamless migration path from shared and switched 10Base-T, to shared 100Base-T, and switched 100Base-T. The IEEE 802.3 100Base-T standard currently includes three separate 100 Mbps specifications:

 

100Base-T4   4 pairs of Cat3 UTP (or better -- Cat4 or Cat5)

100Base-TX   2 pairs of Cat5 UTP, or 2 pairs of DGM Type1 STP

100Base-FX   2 strands of 62.5/125Ám optical fiber

 

   During the evolution of the IEEE's Ethernet standards, the Electronics Industry Association and Telecommunications Industry Association were developing the Commercial Building Telecommunications Cabling Standard (SP2840), formally dubbed EIA/TIA 568-A. This set of cabling specifications defines a generic cable plant capable of supporting a variety of technologies from telephony and data communications to local area networks. Included in the standard are definitions and installation requirements for horizontal and backbone cabling systems and definitions of the physical characteristics for several different media types including UTP, STP, and optical fiber.

   For perhaps the first time in the history of LANs, a set of cabling standards and a set of LAN standards have converged. The maturing of 802.3 to support higher speeds with 100Base-T in parallel with the evolution of cabling technologies and the EIA/TIA standards provides organizations with a simplified set of configuration rules. At least in theory, if one follows the EIA/TIA 568-A standards for their cable plant, that cable plant should be capable of supporting any of the current (and some of the emerging) fast LAN technologies. This support would extend to 100Base-T4, 100Base-TX, 100Base-FX, 100VG AnyLAN, CDDI, and of course 10Base-T, and 4/16 token ring.

   A dynamic and ever-evolving technology, some of Ethernet's latest options under consideration and/or development include 100Base-T2, 1000Base-T, Full Duplex/Flow Control, BLAM, PACE, and FDDQ. <See Sidebar>

   For some time there have been debates over the viability of Ethernet's network access method: Carrier Sense, Multiple Access, with Collision Detection (CSMA/CD). Some network planners do not like Ethernet because of its chaotic, random access approach to channel arbitration and its performance is difficult if not impossible to forecast. Still, Ethernet's popularity continues to grow. According to Dataquest's 1995 Worldwide Intelligent Hub Forecast, over 49 million new Ethernet ports will be sold this year. IDC estimates indicate that the 10 Mbps Ethernet network interface card market continued to grow strongly though out the past year; unit volume rose in 1995 by 27%, totaling 23.7 million units, compared to 18.6 million in 1994.

   When 802.5 token ring was released ten years ago, many thought it would displace Ethernet as the number one choice in LAN technologies. After all, token ring was the first LAN to use anything close to a structured cabling system. It supported a star-wired twisted pair cable plant and provided a deterministic access method that supported access priority levels where Ethernet does not -- never mind that there are few if any applications or protocols that can exploit this Data Link layer priority scheme.

   Additionally, token ring's deterministic access method can only guarantee that all stations will experience network performance degradation equally. Unless applications and protocols are especially written to exploit it, its Token Priority access scheme can only be assigned per physical station. Contrary to the implication, token ring cannot provide a Constant Bit Rate (CBR) which is essential to live, interactive applications such as video teleconferencing. The same is true for 100VG AnyLAN's DPAM and ATM -- as of yet there are no applications that can take advantage of their priority schemes.

   Couple these points with the fact that Token Ring forces every transmitted bit through every potential point of failure in the LAN and it seems that many of the perceived benefits of Token Ring are rendered meaningless. Another drawback to implementing token ring has always been its high cost -- roughly 3 to 5 times that of Ethernet. The costs of Ethernet hardware have been steadily dropping over the years as performance and features have increased thereby further boosting Ethernet's popularity.

   These debates are being revisited along with several new issues now that LANs are moving to speeds of 100 Mbps and higher. For example, whether at 10 or 100 Mbps, Ethernet's channel arbitration is based on signal timing. The 64 byte SlotTime used by the access method restricts the physical extent of the network. This is why the physical extent of a 100Base-T Collision Domain is so much less than that of 10 Mbps Ethernet.

   Channel Capture is another issue that may or may not prove to be problematic. Channel Capture has been identified as a phenomenon that can generate lengthy access delays and cause the back-off algorithm in CSMA/CD to handle contention in a less than fair manner under certain conditions. An alternative back-off algorithm called BLAM (Binary Logarithmic Arbitration Method) has been proposed as a solution. Additional information about Channel Capture and BLAM may be obtained from the following Internet sites:

ftp://ftp.utexas.edu/pub/netinfo/ethernet/techrept13.ps <postscript format>

gopher://mojo.ots.utexas.edu:70/00/netinfo/ethernet/ethernet-capture/usenet/rich-seifert-on-capture-effect

   Interpacket Gap Shrinkage (IPG) can occur in Ethernet LANs consisting of multiple segments interconnected by repeaters. IPG shrinkage can occur because two successive packets may experience differing [Preamble] bit loss on the same path. If this IPG shrinkage is severe enough the two successive packets could be concatenated, or appear to a receiving station be one very long, invalid packet. To a network management system, IPG shrinkage may manifest itself as illegally long frames, giants, or over-runs.

   Micro-segmentation and/or the implementation of full duplex switching at both 10 and 100 Mbps corrects such problems and renders most of these issues moot. For more information on these topics, point your Web browser at:

http://dorado.crpht.lu/CNS/html/PubServ/Network/ether/utexas/papers.html

 

100Base-T vs. 100VG AnyLAN:

   Both 100Base-T and 100VG AnyLAN are technically viable and product pricing is about the same. Academically, 802.12 100VG may appear superior, but as stated previously, no current applications or protocols support DPAM's Data Link layer prioritization. 100VG can, depending on the implementation, support either Ethernet or token ring frame types, but not both concurrently. Like the priority access scheme in token ring, DPAM cannot guarantee a Constant Bit Rate (CBR), so its benefit is limited.

   More vendors support 100Base-T than 100VG. 100Base-T interfaces are available for routers such as Cisco's 7000 series, and Hewlett Packard provides a 100VG-AnyLAN interface card for their AdvanceStack Router 650. To provide a means of interconnection between the two fast LANs, HP has developed a 100VG to 100Base-T bridge.

   Design issues are another point of contention for the fast LAN vendors; 100VG vendors claim that 100Base-T is not backward compatible with 10 Mbps Ethernet. They are referring to compatibility with the 10Base-T cable plant and Ethernet's basic repeater rules. Although it is an insignificant point, they are partially correct.

Some clarifications:

   IEEE 802.3 10Base-T was designed to support two pairs of ordinary DIW-24 phone wire. 802.3 100Base-T(4) and 802.12 100VG were designed to support four pairs of Category 3 UTP or better. 100Base-TX requires two pairs of Category 5 UTP.

   All IEEE 802.3 10 Mbps networks are limited to traversing a maximum of 4 repeaters and 5 segments; this is the maximum diameter of the Collision Domain. 802.3 100Base-T defines two classes of repeater, Class I and Class II. Class I repeaters are restricted to just one repeater in the Collision Domain and Class II repeaters are restricted to 2 repeaters in the Collision Domain with up to 5 meters between repeaters. The diameter of the 100 Mbps Ethernet Collision Domain using UTP is limited to a maximum of 200 meters with a Class I repeater or 205 meters with two Class II repeaters (100Base-TX & 100Base-T4).

   Installation options for optical fiber are a bit more complicated. A single, point-to-point 10Base-FX segment may be up to 412 meters in length. However if 10Base-FX repeaters are used, the distance between any two stations will be less; this is also the maximum diameter of the 10Base-FX Ethernet Collision Domain.

   If a single Class I repeater is used to join fiber link segments, the maximum distance between any two stations is 272 meters. If a single Class II repeater is used to join fiber link segments, the maximum distance between any two stations is 320 meters. Finally, if two Class II repeaters are used to join fiber link segments, the maximum distance between two stations is 228 meters.

   Particularly useful in backbone systems when implemented in full duplex, 100Base-FX will support Link Segments up to 2 km. It is important to note that implementing full duplex mandates the use of switches (or bridges or routers).

   This does alter the design and configuration of 100 Mbps Ethernet relative to 10 Mbps Ethernet, but 100Base-T was designed with switches in mind. By using 100Base-T switches to complement 100Base-T repeater/hubs, the network can be extended far beyond the confines of a single Collision Domain. Switches are more expensive than repeater/hubs, but the per port prices of switches is dropping rapidly. In some cases, the switch price per port is less than that of 10Base-T repeater/hubs just a few years ago! To provision Ethernet bandwidth as needed, several 10 Mbps Ethernet and 100 Mbps Fast Ethernet repeater/hubs can be connected to one 10/100 Mbps Fast Ethernet switch. Stations attached to the 10 Mbps repeater/hubs share the 10 Mbps between all active stations attached to that hub. Stations attached to the 100 Mbps repeater/hubs share the 100 Mbps between all active stations attached to that hub. Individual servers and user stations may also be directly attached to a port on the 10/100 Mbps Fast Ethernet switch, providing those stations with dedicated bandwidth of either 10 or 100 Mbps.

   Keep in mind that for hubs to be attached to a switch port, that switch must support multiple MAC addresses on each port supporting a hub. This is one of the definitions for an enterprise switch. By contrast, a workgroup switch may only support one MAC address per port, restricting that switch to supporting only individual stations on each port (and preventing it from supporting hubs). Most workgroup switches provide at least one up-link port that can support many MAC addresses (usually several thousand). This port is intended for the purpose of attaching to the rest of the enterprise's network.

   Networth Inc., Irving, Texas overcomes the 2-repeater limit through its special Smart Uplink Module (SUM), which is in essence a bridge. By implementing a bridge at the up-link port on their 100Base-T repeater/hub, Networth has bounded the Collision Domain within each hub. Multiple hubs may be cascaded via an external backplane interconnection, thereby creating from one to three logical Collision Domains within the repeater/hub stack.

   100VG AnyLAN vendors often point out that it is possible to construct a single 100VG LAN that is much larger than a single 100Base-T LAN. 100VG vendors also site the ability to support cable runs in excess of 100 meters and the ability to cascade more hubs than 100Base-T will permit.

   While these facts may be accurate, they are irrelevant. First, we have learned that building huge LANs without segmentation is not a good thing -- observe the explosive growth of the switch market. Second, exceeding 100 meters in a cable run violates the EIA/TIA 568-A cabling standards. However, for this issue to be of concern one must acknowledge the inherent value of generic, standardized, structured cable plants.

Two opposing forces:

   The laws of Nature vs. the laws of Man. Physics may support higher transmission rates and greater distances over UTP, and we may have the technology to support it, but in some cases the FCC says the transmission generates and radiates too much EMI/RFI.) Although many signaling methods can achieve 150 meters or more without violating FCC regulations, the EIA/TIA restricts all UTP installations (Cat3, 4, 5) to 100 meters end-to-end.

   The difference between each Category is the spectral bandwidth that can be supported. Greater spectral bandwidth has the potential of supporting faster signaling methods or better data encoding schemes that provide higher bit per second rates. There is no direct correlation between spectral bandwidth (measured in MHz) and data throughput (measured in Mbps).

<< CHART: EIA/TIA 568-A media >>

   Don't bother with coax. Existing coax-based LANs do require maintenance and may perhaps require some expansion, but coax should be viewed as a declining technology. None of the new and emerging fast LAN technologies support coax. Any company or consultant that recommends coax as the media of choice for use in a new installation is doing themselves or their customer a disservice. There simply is no future in coax LANs.

A reality check:

   The installation requirements for Cat5 are so stringent that perhaps as many as 20% of all Cat5 cable runs are NOT Cat5 compliant. Further, challenges to the reliability of the test and certification procedures for Cat5 installations are still being addressed and the standards for testing Cat5 have yet to be completed. Still, to provide support for tomorrow's technologies as well as today's, it is advisable to design all new cable plants such that they adhere to the EIA/TIA 568-A standards.

Alternatives exist:

   As mentioned previously, IEEE 802.3 100Base-T4 (and 802.12 100VG-AnyLAN) supports Cat3 or better. There is apparently no shortage of Cat3. Cat4 has been eclipsed by Cat5, the cost difference between the two is negligible and very few companies still manufacture it. Additionally, the cost and complexity of installing fiber to the desktop has dropped dramatically -- now is perhaps a good time to consider fiber to the desktop.

   Plan, design, and install for the future. It only makes good business and financial sense to plan on installing a new cable plant that will satisfy an organization's networking requirements well into the foreseeable future. Do not skimp here. The cable plant is the foundation of any network. If the foundation of your network infrastructure is not stable, everything built on top of it will be unreliable.

In summary:

   802.3/Ethernet is alive, well, and growing. Category 5 UTP and optical fiber are in, coax is out. LAN and cabling standards are converging to provide generic support for any of the new fast LAN technologies as well as some of the older LAN technologies. Switching can provide greater bandwidth availability where needed and Fast Ethernet provides greater bandwidth where needed. 100Base-T Fast Ethernet support and access can be provided via repeater/hubs, switches, and routers from multiple vendors. Used in concert, 10/100 shared and switched Ethernet over copper and/or fiber can be tailored to provide an extremely flexible and cost effective means of networking any enterprise.

   Design cable plants to be EIA/TIA 568-A compliant, but consider exceeding the basic media requirements. Rather than running one jacket of 4 pair Cat3 and one jacket of 4 pair Cat5 to each work area as the standard requires, run three jackets of 4 pair Cat5 to each work area. To further future-proof the cable plant, add 8-strand optical fiber to each work area. The cost of the additional materials is relatively minor. The cost of recabling can often be prohibitive and the disruption of productivity during recabling unacceptable. Therefore, whether planning to recable or a new installation, plan for the future, do it once, and do it right the first time.

 

Sidebar (Some new stuff in the works)

802.3w - BLAM

    The Binary Logarithmic Access Method is intended as a potential alternative to Ethernet's existing truncated Binary Exponential Backoff algorithm (BEB). Rather than replacing the current system, BLAM would coexist with BEB to provide backward compatibility to legacy LAN systems. Observations of certain adverse effects of Ethernet behavior on networks using high-performance workstations and servers was the driving force behind the development of this alternative algorithm. [The use of Ethernet LAN switches can negate these adverse effects.]

802.3x - Full Duplex/Flow Control

   Flow Control will allow vendors to build switches with finite, limited memory, and to prevent packet loss by issuing flow control commands to the attached stations. Pushing the problem back to the station is much preferable to throwing away the packet in the switch, due to buffer overflow.

   The problem is especially bad when using full duplex, as there is no way to "back-pressure" the station by asserting false carrier or false collisions, since the station ignores these in full duplex. If (for example) you have a lot of workstation ports all sending traffic to a single server port, there is nothing the switch can do to prevent getting congested. The idea of flow control is to stop the senders from sending more traffic when this condition occurs, rather than just discard frames. Secondarily, there is a desire to build switches with very limited memory, for cost reasons. This also requires more flow control.

   Flow control doesn't change CSMA/CD, rather it is in addition to it. The plan is to implement flow control using standard, valid frames (with a reserved Type field for the purpose). There is no change to the underlying access protocol. There is no impact on the low-level hardware. The older ESPEC Type field format was chosen to avoid having switches implement 802.2 Logical Link Control layer functions. The IEEE 802.3x sub-committee is actually considering a proposal that would allow the older DIX ESPEC v.2 Type fields in a general manner for use in 802.3 LANs!

802.3y - 100BASE-T2

   A new 100 Mbps signaling scheme that should support 100 Mb/s operation over two pairs of Category 3 UTP. Assuming that it works as advertised, it will support 100 meter cable lengths and is capable of full duplex operation.

   The current proposal uses an encoding scheme called "PAM 5x5". It is 5 level, 5 phase (hence 5x5) constellation-space signaling method. 100Base-T2 is being run in "dual duplex," i.e., both pairs are used in both directions simultaneously (and use hybrid cancellers to remove the transmitted signal from a received signal).

   The system is extremely complex and would not have been practical even a few years ago. However, with the availability of low cost, 0.6 micron silicon technology, the large amount of circuitry required for the complex digital signal processing is possible "in a corner of the chip".

802.3z

   Gigabit Ethernet -- 1000Base-T [Just when we thought the 802.3 committee would have to stop work because they have run out of letters... they now use double letters such as 802.3ab which is the committee working on Gigabit Ethernet over UTP.]

   This is dependent upon acceptance of the PAR (Project Authorization Request) by the 802 Executive Committee at the March Plenary. The main thrust of the Gigabit Ethernet work is using fiber. Specifically, the plan is to run Ethernet over Fiber Channel, similarly to how 100Base-T adopted the ANSI X3T9.5 FDDI TP-PMD (aka., CDDI) specification to support 100 Mbps operations over UTP. There is some interest in running Gigabit Ethernet on multiple pairs of Cat5 UTP, over short distances of perhaps 50 meters. Currently, a key interest is in full duplex switch-to-switch connections for backbone applications.

FDDQ: Fair Dual Distributed Queuing

   At this time, FDDQ is just a concept, it is not part of any IEEE working group. It would supposedly provide utilization and average latency comparable with those of CSMA/CD, but puts packets into one of two global distributed queues when the network becomes congested. In this way, FDDQ would provide two-priority access levels to the network.

PACE: Priority Access Control Enabled

   Developed by 3Com Corporation, PACE is not part of any IEEE working group. It is a switch-based technology that makes use of installed PCs, workstations, Ethernet adapters, cabling and management tools, and preserves the investments in administrative and management expertise. PACE technology addresses the problem of variable network delay through a variety of techniques to regulate Ethernet timing and deliver high-quality, real-time multimedia.

   An Ethernet switch with PACE employs traffic control algorithms that allow each link to the switch to operate at more than 98% efficiency, even under full load and when servicing a mix of real-time and conventional data traffic. The traffic control algorithms also provide predictable LAN transmission on Ethernet. PACE technology uses star-wired switching configurations and enhanced Ethernet. The same enhancements work at both 10 Mbps and the Fast Ethernet speed of 100 Mbps.

   3Com's PACE technology is supported by its strategic partners which include Apple, Dell, Novell, Oracle, Silicon Graphics, Starlight Networks, and Sun Microsystems. 3Com licenses the technology to other hardware vendors.

[Special thanks to Rich Seifert, President, Networks and Communications Consulting for his assistance in compiling and sorting through much of this standard's information. Mr. Seifert is also the chair of the IEEE 802.3x committee, however the opinions and comments mentioned herein are Mr. Seifert's and my own.]

 

Market Stats

Ethernet The 10Mbps Ethernet network interface card market continued to grow strongly; unit volume rose 27 percent, totaling 23.7 million units, compared to 18.6 million units in 1994. Revenue topped $2.0 billion from $1.8 billion in 1994. As vendors sought to preserve market share, the familiar pattern of price cuts continued during 1995 but at a somewhat gentler pace on average than during 1994. By year-end, however, a few models of Ethernet cards had slipped below the $60 street price threshold in the United States. 3Com continues to dominate Ethernet unit volume and increased market share during 1995.

Token Ring Token Ring unit shipments, meanwhile, totaled 3.8 million units, up almost 19 percent from 3.2 million units in 1994 and slightly above expectations. Token Ring revenue registered just under $1 billion, an anticipated 25 percent drop from almost $1.3 billion in 1994.

High-Speed LANs Although high-speed LAN adapters (including ATM, 10/100 Base T, and VG AnyLAN) had an insignificant impact on the overall market in 1995, sales are growing rapidly from a small base. Total units shipped approached one million for the year. The 100 Mbps products are starting to ramp up as more vendors jump into the game. IDC believes the sales outlook is improving with the delivery of a range of 100Mbps hubs.

LAN Switching The 1995 LAN switch market grew at a dramatic rate from that of 1994; total port shipments for 1995 surpassed 2 million ports, representing a growth rate of 390 percent over 1994 port shipments of 413,000. Revenue grew by more than 300 percent from $350 million in 1994 to nearly $1.5 billion in 1995.

Routers Despite mounting pressure from LAN switches and the saturation of customer networks, routers continue their strong growth in unit shipments and revenues. For 1995, router unit shipments grew to approximately 562,000 units - up 71 percent from 1994's total of 328,000 unit shipments. Revenue also grew, but at a more conservative rate (42 percent) - reaching $3.7 billion for 1995 versus $2.6 billion for 1994.