Gigabit/2: give your network connectivity a performance boost

By Jonas Buys

1. A short Introduction to Gigabit.
At all levels of the network, the need for speed continues to accelerate, and today's servers and PCs are designed to meet the challenge - requiring the right network connections. High-bandwidth applications and high-speed communications have driven major increases in networking speeds from the prehistoric 10Mbps Ethernet to the current widely in use 100Mbps Fast Ethernet, and now is evolving to reach a standard of 1000Mbps Gigabit Ethernet. A current trend that ever emerging is that many IT managers are deciding the time is ripe for broadly implementing Gigabit Ethernet in existing networks. They are specifying 1000Mbps network adapter cards or LOMs (LAN-on-Motherboard) for their desktop systems, and multi-Gigabit Ethernet for their servers. Benefits to the IT manager include significant time savings for a myriad of daily tasks ranging from virus protection patches to backups and software pushes. Users on the network gain the ability to download seamlessly without waiting, greatly increasing their productivity.

This article looks at the ways your network can benefit from Gigabit implementations and offers a variety of deployment examples using copper and/or fiber-optic cabling. This information is intended to help network managers stay ahead of the demand for bandwidth and get the most performance at a lower cost, in their existing OS/2 networks.

A recent study from International Data Corporation (IDC, 2000) showed out that 87% of all installed network connections are Ethernet-based. Primarily, this is due to the fact that industry standards for Ethernet, of which some are already more than 25 years old, have progressed along with networking requirements. This progression of industry standards provides a clear and easy migration path for companies as their bandwidth requirements increase.

Gigabit Ethernet has evolved from the original industry standards for 10Mbps Ethernet (10BASE-T) and 100Mbps Fast Ethernet (100BASE-TX and 100BASE-FX). When the IEEE (Institute of Electrical and Electronics Engineers, Inc.) approved Gigabit Ethernet over fiber-optic cable in June 1998, companies were able to rely on a well-known, standards-based approach for improving traffic flow in congested areas. Gigabit Ethernet began to be used along network backbones and in network servers to remove traffic bottlenecks in these areas of congestion. As Internet-based activities increased, companies were also deploying Gigabit links from workgroups into the data center. To implement Gigabit Ethernet, however, network managers were often challenged to re-wiring their buildings in order to upgrade the infrastructure to fiber-optic cable. This issue was resolved in 1999, when the IEEE adopted an industry standard for Gigabit Ethernet over classic Category-5 copper cabling. Hence, widespread deployment of Gigabit Ethernet became possible and more affordable over the existing copper infrastructure.

The revolution in productivity that began with the advent of the desktop computer still hasn't reached its end. This revolution continues today with the rapid growth of the Internet, which requires desktop PCs to quickly download large files, process rich content, and communicate as well as compute. Not long ago, desktop connections and PC processing power were considered more than adequate if the user could simultaneously open a word processor and a spreadsheet application. Some wondered if the 10Mbps of bandwidth would ever be fully utilized. The kind of visually rich multi-tasking that is routine today, such as preparing a presentation, embedding graphs created with your spreadsheet in it, while downloading photographs or exchanging graphic designs while talking on an Internet phone line via VoIP, were not part of the equation. Then, as applications grew more bandwidth-hungry, the view that 10Mbps was adequate for desktops began to change, and the trend to 100Mbps desktops developed. This trend greatly accelerated when the cost of dual-speed 10/100Mbps Ethernet connections approached the cost of regular Ethernet. Network managers could build 100Mbps capability into their new PCs at the time of purchase and avoid the higher costs of retrofitting these desktops in the future. The industry is now in a similar circumstance with 1000Mbps Gigabit Ethernet connections.

Behind the growth in bandwidth requirements has been a revolutionary change in the way people get their work done. Several years ago, people relied on the network mainly to deliver productivity applications such as file and print. An enormous shift has occurred with the advent of the Internet, intranets, and the widespread use of workgroup applications such as Lotus Domino and Notes, and WebSphere Application servers. IT managers have redesigned the network to accommodate this shift, and most networks now trust their traffic to the robust and reliable TCP/IP. Desktops are being used to download massive amounts of data from the Internet and intranets, and networks are far busier, since collaboration and simultaneous access to databases is ever increasing.

2. Chipsets and Drivers.
A diverse set of Gigabit chipsets is currently available to the market, and as development of the technology continues, newer ones are released on regular basis. Latest Gigabit chipsets already support speed up to 10Gbps! To really take advantage of such Gigabit NICs, the traditional parallel PCI architecture doesn't suffice anymore. More information about new PCI technologies and Gigabit NICs can be found in the PCI-X section of this web site. In what follows, we'll have a brief description of all Gigabit solutions available for the OS/2 platform family.

2.1. 3Com 3C2000-T Gigabit Ethernet (and Syskonnect SK98xx).


Being a classic designer of network interface cards (NICs), 3Com Corporation has a select but powerful series of Gigabit products available, both targeted at the server market (Fiber-optic wires) and desktop market (copper). Unfortunately, none of these cards is officially supported on OS/2 by 3Com.

There is one exception, though. The 3Com 3C2000 Gigabit NIC can be used on OS/2 and eComStation using Willibald Meyer's GenMAC device driver, that also support other modern fast ethernet and gigabit NICs. This device has 0x10B7 and 0x1700 as Vendor and Device IDs respectively.

In fact, GenMAC supports the Marvell Gbit Core IP chipset, which is used in other Gigabit NICs too, like the Syskonnect (Schneider &amp; Koch) SK98xx series.

2.1. Broadcom Solutions 57xx NetXtreme Gigabit Ethernet.


As the leading provider of Gigabit Ethernet chips, Broadcom offers a field-proven family of Gigabit Ethernet controllers that are generally considered as the best of all Gigabit solutions available today. Broadcom has licensed out its chipsets to Hewlett-Packard (Compaq), IBM, DELL and some other companies, which now ship their own-branded Gigabit network interface cards (NICs) in their high-performance servers. Most of the time, these companies officially support OS/2 Warp Server for e-Business for use as server platform with these servers.

Broadcom has developed and officially supports an OS/2 driver that offers support for all chipsets that belong to the 57xx chip family. In particular, this driver supports the BCM5700, BCM5701, BCM5702, BCM5703, BCM5705, BCM5721 and BCM5751 chipsets. The driver supports all network interface cards based on this chipset family, as well as branded NICs of IBM, HP and others. The driver can also be used to make on-board gigabit solutions work.

The Broadcom 57xx Controllers can be used both on PCI and PCI-X systems, and the OS/2 drivers have been designed to support an implementation with PCI-X too.

The Broadcom VendorID is 0x14e4, and the DeviceID of the chipsets can vary: 0x1644, 0x1645, 0x1646, 0x1647, 0x165d, 0x165e are very current, and for HP devices the DeviceID is 0x1696.

2.2. Intel Solutions.


Leading chip-baker Intel offers a broad range of Gigabit network cards based on their own chipsets for both copper as fiber connections. Intel has a wide range of Gigabit controller chipsets, and targets different products to the server and desktop markets.

Nearly all products of the Intel PRO/1000 family are supported, including the DualPort models. QuadPort products are not supported though. Intel offers official OS/2 support for their Server-targeted gigabit solutions, but also has drivers available for their Desktop-targeted products. With the drivers Intel offers a very decent and performant solution to their corporate customers. The Intel drivers can be downloaded free of charge from www.intel.com. Just download PRODOS.EXE and PRODOS2.EXE.

Additionally, IBM comes with a device driver through SoftWare Choice for a selected set of Intel chipsets. At the right you can find a table with the different chipsets used with Intel PRO/1000 Gigabit NICs. All Intel-related device drivers support both Copper and Fiber wiring.

Included in the table above are the Vendor and Device IDs you can use with pci.exe utility to see which chipset your Gigabit Intel card has.

2.3. TMI/Tamarack TC9021 32/64bit PCI 10/100/1000Mbps Ethernet MAC Controller.


TMI/Tamarack have recently been taken over by the company [IC Plus], rather a strange bird for networking products. The old TMI products are still being produced and dispatched, and a driver for one of their gigabit chipsets has been developed by Nobuyuki Yanagihara and is available at [os2warp.be]. You can get support from the developer via [os2warp.be Technical Support Center].

The TMI TC9020/21 chipsets have the 0x13f0 (Sundance) VendorID and 0x1021 as DeviceID (though 0x143d and 0x9021 are also possible values respectively).

This chipset is rarely used in NICs, and it doesn't offer all the features the other chipsets come with. This device driver will only work with Copper wiring.

2.4. National Semiconductor DP83820 10/100/1000MBit Gigabit Ethernet.


The DP83820 Gigabit chip from [National Semiconductor] is targeted at high-performance adapter cards and mother boards. The DP83820 fully implements the V2.2 66 MHz, 64-bit PCI bus interface for host communications with power management support. Packet descriptors and data are transferred via bus-mastering, reducing the burden on the host CPU. The DP83820 can support full duplex 10/100/1000 Mbps transmission and reception.

A driver for this chipset has been developed by Nobuyuki Yanagihara and is available at [os2warp.be].

Notice their are two different drivers. The &quot;modified&quot; driver is the oldest and is no longer being developed. These drivers are based on the source code of the DOS driver that was released by National Semiconductors*. However, we strongly discommend the use of this driver and recommend the use of the other driver since it has been written from scratch and delivers far better performance.

Another driver that supports the NSC 83820 chipset and that is very performant is Willibald Meyer's GenMAC driver (also available via the Gigabit Drivers Page), which also supports a lot of other chipsets and modern network cards. You can get support from the developers of both drivers via [os2warp.be Technical Support Center].

The NSC DP83820/21 Gigabit chipsets have VendorID 0x100b and DeviceID 0x0022. This device driver will only work with Copper wiring.


 * Please note these are NOT an official National Semiconductors device drivers.

2.5. Realtek RTL8169 10/100/1000M Gigabit Ethernet.


The Realtek RTL8169 chipsets family consists of 2 related gigabit controller chips, in fact: the RTL8169S-32 and the RTL8169S-64. With these robust chips, [Realtek] gives the world an answer to Gigabit. Just as Broadcom, Realtek licenses its controller to other companies that can then produce their own-branded Gigabit NICs with it. However, not a lot of companies have licensed the Realtek chipset, since it is not as perfect as the Broadcom chip is.

The devices support the PCI v2.2 bus interface for host communications with power management and are compliant with the IEEE 802.3 specification for 10/100Mbps Ethernet and the IEEE 802.3ab specification for 1000Mbps Ethernet. The devices support an auxiliary power auto-detect function, and will auto-configure related bits of the PCI power management registers in PCI configuration space.

The RTL8169 comes with all standards implemented that you could ever wish for, but the OS/2 driver does not support all of them at this moment yet.

A driver for this chipset has been developed by Nobuyuki Yanagihara and is available at [os2warp.be], and also Willibald Meyer's GenMAC driver supports it. You can also get support from the developer via [os2warp.be Technical Support Center]. The Realtek RTL8169 Gigabit chipset has 0x10ec and 0x8169 as Vendor and Device IDs, respectively. This device driver should support both Copper and Fiber wiring (Copper has been tested successfully).

3. List of OS/2 CHL Tested Devices.
The following devices have been tested by OS/2 CHL team and have been found 100% compatible with OS/2 and eCS. For the most up-to-date listing of tested devices (this is only subset of those in the list), please consult the [OS/2 CHL]], where you'll also find the appropriate chipsets used in each device listed below and the drivers to use.

This listing of devices may not be copied, published elsewhere nor edited without explicit written permission of the author.

4. Wiring the gigaconnections: fiber or copper?
A study dated from 2000 by Sage Research concluded that 87% of companies are running their networks on Cat-5 cable, and the report showed that approximately 90% of these existing networks should be able to run Gigabit Ethernet over the existing copper wiring. ( Legacy Cat-5 cable destined for 1000BASE-T use should be tested for far-end crosstalk and return loss, and corrected if necessary. If the cabling link doesn't pass, ANSTI/TIA/EIA industry-standard TSB-95 (1998) defines five relatively simple options for correcting performance.)

For new cable installations, network designers might want to consider the enhanced Cat-5 cable (Cat 5e) to gain extra signal margin. However, this is not a requirement for Gigabit throughput. Network managers deploying Gigabit Ethernet have a choice of copper or fiber to match different situations. For example, fiber is typically reserved for situations that require cabling distances greater than the 100-meter copper limit - such as between buildings, or for physically sizeable networks. Copper is most widely deployed for horizontal applications in walls and ceilings. Both copper and fiber are used in vertical risers between floors throughout office buildings. In the next section, some attention will be spent to the basic theories behind both cabling possibilities, and afterwards, a better look will be spent at the advantages of fiber wiring.

4.1.2. Fiber Optic.
Fiber optic refers to the medium and the technology associated with the transmission of information as light impulses along a glass or plastic wire or fiber. Fiber optic wire carries much more information than conventional copper wire and is far less subject to electromagnetic interference. Most telephone company long-distance lines are now fiber optic. Transmission on fiber optic wire requires repeating at distance intervals. The glass fiber requires more protection within an outer cable than copper. For these reasons and because the installation of any new wiring is labor-intensive, few communities yet have fiber optic wires or cables from the phone company's branch office to local customers.

Data is sent over fiber optic cables by a Physics process known as Total Internal Reflection.



Consider the image at the right below. A basic knowledge of Optics is assumed. When light passes from a medium with one index of refraction m1 to another medium with a lower index of refraction m2, it bends or refracts away from an imaginary line perpendicular to the surface (normal line). As the angle of the beam through m1 becomes greater with respect to the normal line, the refracted light through m2 bends further away from the line. At the so-called critical angle, the refracted light will not go into m2, but instead will travel along the surface between the two media (sin[critical angle]=n2/n1, where n1 and n2 are the indices of refraction (n1 &lt; n2)). If the beam through m1 is greater than the critical angle, then the refracted beam will be reflected entirely back into m1 (total internal reflection). In Physics, the critical angle is described with respect to the normal line. In fiber optics, the critical angle is described with respect to the parallel axis running down the middle of the fiber. Therefore, the fiber-optic critical angle = (90 degrees - physics critical angle). In an optical fiber, the light travels through the core (m1, high index of refraction) by constantly reflecting from the cladding (m2, lower index of refraction) because the angle of the light is always greater than the critical angle. Light reflects from the cladding no matter what angle the fiber itself gets bent at, even if it's a full circle! Because the cladding does not absorb any light from the core, the light wave can travel great distances. (However, some of the light signal degrades within the fiber, mostly due to impurities in the glass.)

4.2.1. Bandwidth & Speed
In today's high transfer environment, where speed and low signal loss are critical parameters, it is increasingly evident that copper wires will be inadequate for most future applications. Copper's major drawback is that its transmission capability is only 100 Mbps for up to approximately 100 meters. This is due to the distributed capacitance which is an inherent problem in all copper cable. (i.e., The greater the distance, the more capacitance until after 100 meters, the signal is no longer adequate.) This means that for any large Local Area Networks wired for copper, such as might be found in an average corporate complex, a series of expensive repeaters must be installed. Each of these repeaters will also require a corresponding number of wire closets, all taking up valuable real estate. As a result, servicing a malfunctioning device on a copper based network becomes very complicated, requiring engineers at the switch station, workstation and wiring closet.

Fiber optic cable has no such limitations. The average bandwidth for multimode fiber is 500 MHz for one kilometer. In a standard office environment, this would eliminate the need for repeaters and extra closet space. The concentrators could be run to a central location in a topography known as &quot;homerunning&quot; and fewer personnel are needed to diagnose and correct a problem. Furthermore, because of fiber optic cable's greater transfer capabilities, the need for the replacement of a system will not be as frequent as it is for a LAN wired with copper. For example, a currently installed shielded twisted pair system with an expected lifespan of eight years will have to be re-cabled after only 3-4 years due to the increased bandwidth requirements. Fiber optic cable, on the other hand, is currently ready to accept protocols and interfaces such as FDDI (Fiber Distributed Data Interface), ATM (Asynchronous Transfer Mode), ESCON (Enterprise Systems Connection), Fibre Channel and perhaps any additional mode that may arrive or be proposed in the near future. Because of these capabilities it is believed that the lifespan of a fiber based LAN with multimode fiber is at least fifteen years.

Apologists for copper like to point out that copper has advantages over fiber in two areas. The first supposed advantage is that copper is less expensive and less complicated than fiber to install. The second advantage is that copper can withstand greater punishment in the horizontal section. Both of these claims are relative, and for each situation you should should consider the pro and the cons of either copper and fiber.

4.2.2. Cost
Fiber optic cable does have a premium over copper initially, particularly during installation. But as has been addressed above, the greater transfer rate, permanence and ease of maintenance causes these costs to drop in the long term. Because fiber is lighter and more flexible than copper, it is also easier to install. All in all, for large networks, fiber optic cable is actually less expensive overall than copper. While at one time fiber installation needs at least 15 minutes and the ability of a highly skilled technician to prepare the connectors, today's simple crimping technology and fast curing epoxies have cut that time to five minutes or less. Couple this with the increased time it takes to install connectors and the labor costs for fiber installation versus copper are now roughly equal. Most copper cost estimates also do not take into account the more involved wiring schematics (i.e., wiring closets, repeaters) needed for UTP installation. Perhaps the most important cost issue, and one rarely examined in the fiber versus copper debate is copper's imminent obsolescence. As mentioned before, due to the increased signal traffic that LAN's will be expected to carry in the near future, copper's useful lifespan is limited to at most four years. After this, it will be inadequate for most serious needs, and will have to be replaced. This replacement will, of course, entail the large scale dismantling of not only the copper itself, but also of all the extensive support equipment. When all is said and done, the short term cost advantage of installing copper over fiber will not only have evaporated, it will have become a huge cost deficit.

4.2.3. Size, weight and some other things...
A fiber optic cable with the same information carrying capacity (bandwidth) as a comparable copper cable is less than 2% of both the size and weight. Large capacity copper cables lead to overcrowded chases and conduits. They take up a lot of space and require special support structures to handle the weight. Even when the individual cables are not large, the aggregate can be unwieldy and difficult to manage. Consider a typical copper patch panel environment which may terminate 200 four-pair STP cables, each weighing 50 lbs. per 1000 feet. A duplex fiber-optic cable weighs 12 lbs. per 1000 feet and is much more flexible - plus the bandwidth is much larger. Fiber optic cables can be installed in the same conduits as power cables. Copper communication cables require separate conduits, or you'll experience a severe damage of the data sent over the network.

4.2.4. Security
Finally, fiber is a secure medium, immune from tapping, as by its very nature (optical characteristics). If the signal is tapped, light loss is unavoidable, and the connection is shut down.

5. And what about Linux?


The Linux Community is following a similar scenario that the OS/2 one: both Broadcom and Intel provide official support and drivers for Linux, and development of other Open Source drivers is also continuing. One possible reference of Open Source drivers is , a publication of Scyld Computing Corporation. An official Linux driver for the TC9020 chipset is available from the former TMI/Tamarack, 3COM officially supports the 3C2000T, and all Syskonnect gigabit NICs come with Linux drivers shipped. In meantime, Realtek has also brought out a Linux driver for the RTL8169 series of chipsets and undoubtedly, many more drivers will be in development behind the scenes.

6. Conclusion.
Deployment of bandwidth-hungry applications and more powerful processors is likely to continue for the foreseeable future, requiring faster network connections. Gigabit Ethernet is a viable technology that allows Ethernet to scale from 10/100 Mbps at the desktop to 100 Mbps up the riser to 1000 Mbps in the data center. Whether you ought to switch to gigabit instantaneously depends on your particular needs, but in general gigabit only shows its benefits on large corporate local area networks and is overkill for just a simple Internet connection. With the eye on the future, deploying Gigabit for application servers, data servers and gateways would be a justified investment. With a select but performant set of copper as well as fiber gigabit solutions available, we have in fact good support for the OS/2 and eComStation platform family. However, the latest extreme-fast Gigabit technologies are not available for use on OS/2 yet. To take real advantage of Gigabit networks, you'll need fiber optic cabling to cover a physically large network. Fiber optic is - compared to costs that arise with regular copper - not significantly more expensive.