ed Loop (FC-AL). Each one has something to recommend it, but they aren't interchangeable.
I won't keep you in suspense here: A syst
em you buy in two years is likely to have a couple of these buses. USB is really too slow for much more than keyboards, mice, printers, and modems. FireWire is faster, but it's best suited for digital peripherals (e.g., cameras, VCRs, and the like). SSA is plenty fast, but it's shaping up to be something you find only in IBM shops.
In the end, it's likely that a form of FC-AL will be your high-speed storage connection. What follows is an explanation of how I came to those conclusions.
Why Serial?
Blame SCSI's problems on the PCI bus. Until PCI arrived on the scene, SCSI seemed impressively fast. But compared to a 33-MHz PCI bus with a raw transfer rate of 132 MBps or a 66-MHz PCI bus running at 524 MBps, SCSI-1 and SCSI-2's measly 5 MBps pales.
So why not just make SCSI faster? That way you could keep at least some level of compatibility. No problem: You can already get 80 MBps from an Adaptec dual-channel ultrawide SCSI adapter. And the ANSI XT310 committee is discussing Ultr
a-SCSI-2 at 80 MBps and Ultra-SCSI-3 at 160 MBps.
But sooner or later you start encroaching on certain laws of physics. Here's why: There are two ways to increase the transfer rate of an interface. You can add more links or speed up the existing links. But both have problems. If you add more wires (and you'd have to add
lots
more), you'd create an unmanageably thick cable. (A standard SCSI cable already has 50 wires, and a wide SCSI cable has 68.) If you speed up the link, on the other hand, you'd make
clocking
-- deciding where one bit ends and the next starts -- increasingly difficult for devices on the bus. The acceptable margin of error in clocking becomes smaller as the interface gets faster and the bits get shorter. When multiple lines have to be clocked in parallel, a mismatch is also possible between lines. Making SCSI faster just seems like more trouble than it's worth.
Recognizing that any radically different approach would take years to get off the ground, engineers began
looking for other approaches that had more growth potential than SCSI. The answer seemed to be serial. Serial interfaces transfer just one bit at a time on just one line. (A full-duplex serial interface transfers one bit at a time in each direction, but on two separate lines.)
In addition, serial interfaces eliminate the possibility of a mismatch between lines, because the clock information is in the same stream as the data. Also, serial interfaces offer simple wiring between the peripheral and the computer: just two wires per line. Therefore, a full-duplex serial interface requires only a four-wire cable.
So, in summary, serial can go fast, and the cables can be thin. But once the engineers got that far, they started to take some radically different tacks.
USB: Parallel Replacement
Although slower than SCSI, the 12-Mbps (1.5-MBps) USB interface offers both higher speeds and Plug and Play capabilities for peripherals that previously would have used the serial or parallel port
or a special port, such as the one for a keyboard, mouse, or monitor. USB can also be used for a small number of peripherals, such as scanners, that in the past used SCSI or proprietary adapters.
For non-SCSI peripherals, USB offers performance that's over 100 times faster than the 115-Kbps maximum of the traditional serial port, for instance. The parallel port generally has a maximum rate of about 2 Mbps, about six times slower than USB.
In addition, up to 126 devices can be daisy-chained through USB hubs to a single USB port -- without requiring any added computer resources, such as IRQs. In contrast, a serial, parallel, keyboard, or monitor port typically supports just one device. Moreover, low-power USB devices can get electrical power through the bus, reducing the need for wall outlets or power-distribution bars. (High-power devices, such as scanners, still need an outlet.)
USB also permits
hot swapping
-- the ability to remove and replace devices without the need to turn off the PC.
Devices such as keyboards and mice, which feature two-way communications, cannot be reliably hot-swapped today. Finally, there is one standard cable for all USB devices, making cable replacement easier and less expensive.
However, the first USB chips became available from Intel and National Semiconductor only in late 1996. Some chips, such as USB hub controllers, won't be available in volume until the second quarter of this year.
The availability of chips is certainly a major factor in the rollout of USB products. Equally important, though, is availability from Microsoft of
class drivers
-- software modules that offer support in Windows for each type of USB device. Microsoft has begun releasing class drivers for Windows 95, the first being a driver for digital cameras. Class drivers for Windows NT aren't expected until release 5.0, which isn't expected until late this year.
USB is
designed for low- and medium-speed devices, such as monitors, keyboards, printers, mice
, modems, and scanners. As far as storage devices go, it can be used for those that connect to parallel ports today, such as removable cartridge and tape drives. However, USB is simply not designed for high-speed devices or applications. That's where SSA, Fibre Channel-Arbitrated Loop (FC-AL), and FireWire come in.
SSA: IBM's Storage Entry
As recently as mid-1995, it looked as if IBM would succeed in promoting SSA as an alternative to SCSI. IBM's main selling points for SSA over SCSI were speed and cabling: SSA offers two full-duplex channels, each running at 20 MBps in each direction, for a total maximum transfer rate of 80 MBps -- double that of Ultra-SCSI. And the SSA cable has just four wires -- two twisted pairs, one pair for each channel.
By 1995, an ANSI committee (X3T10.1) had been working on a formal SSA standard for two years, and the SSA Industry Association boasted more than 40 member companies. Chips were available from IBM. Pathlight Technology (Ithaca, NY) had SSA a
dapters for OEMs. Conner Peripherals (San Jose, CA) was on-board to support the interface in its disk drives.
However,
SSA's promise
as an OEM technology was never fulfilled. "The bottom line on SSA," says Allan, "was that after two years of hoped-for big OEM wins, nobody ever got one. SSA offered no extra bang per buck over SCSI when like systems were compared."
The problem was with the number of DMA lines required. The 80-MBps transfer rate for SSA, notes Allan, assumes that there's data all the time on four 20-MBps DMA lines. "If you hooked up four SCSI adapters (which would take four DMAs), you get 160 MBps, which leaves SSA in the dust," he explains. "And a four-DMA SSA card is more expensive than four SCSI adapters."
In October 1995, Seagate acquired Conner. Seagate, which is a partisan of FC-AL, then quietly buried SSA. (The official announcement that Seagate was discontinuing development on SSA came in February 1996.) Faced with a lack of support from the world's large
st drive manufacturer, other players, such as drive manufacturer Micropolis, also backed away from SSA. Adaptec (Milpitas, CA), seeing that SSA was emerging as an IBM-only technology, shelved its plans to support the interface.
SSA has technical limitations, too. Hybinette, a Boulder, Colorado-based integrator specializing in assembling high-performance Alpha-based servers for the graphics industry, intended to release a line of shared SSA disk arrays with PCI controller cards. However, when trying to share the disk arrays, they had file-corruption problems, and the only cures the company could find were a $3000-a-seat software package or a hardware-based switching scheme. Both were much too expensive, says VP of Sales Bob Simpson. Furthermore, Hybinette found it could achieve only 5 MBps when dedicating an SSA disk to its Alpha-based server.
IBM now offers several four-port SSA adapters for the RS/6000, as well as two high-performance, high-capacity disk subsystems, the 7131 SSA Multi-Storage Tower a
nd the 7133 SSA Disk Subsystem. IBM has also signed OEM deals with La Cie and Groupe Bull.
IBM is still solidly behind SSA. It's even planning a next generation, due to appear in 1998, that doubles link speeds for a total throughput of 160 MBps, according to Bill Pinkerton, director of marketing for open systems storage in IBM's Storage Systems Division in San Jose, California. However, SSA is a technology with a limited lifetime: It will be subsumed in the next generation of Fibre Channel.
FC-AL: 100 MBps and Up
FC-AL
offers a substantial bandwidth improvement over SCSI. FC-AL's base speed is 100 MBps. However, a large number of products are dual-ported designs offer-ing a total bandwidth of 200 MBps. Dual porting also increases fault tolerance by offering redundant access to each device. Base speeds of 200, 400, and 800 MBps and higher are expected during the coming years. In short, FC-AL is starting out, and will probably always be, more than twice as fas
t as the fastest SCSI available.
As if that's not enough, there are three more things to make you put FC-AL on your early Christmas shopping list. First, FC-AL is self-configuring and enables you to hot-plug devices. Second, many host computers already support FC-AL for host-to-host interconnection. Third, FC-AL supports cables up to 30 meters long with coaxial cable, or 10 kilometers with fiber-optic cable. And there are already twisted-pair products available, such as the FibreNet line of cards and disk arrays from Transoft (Santa Barbara, CA) that support runs up to 65 feet.
FC-AL supports the major peripheral command interfaces, including SCSI-3, High Performance Peripheral Interface (HIPPI), and Intelligent Peripheral Interface (IPI-3), minimizing the need for software changes. The only standard command set supported by SCSI is SCSI itself. Also, you can connect an amazing 126 devices to a single FC-AL port; a single SCSI bus can support only 15 devices.
FC-AL primarily targets high-end serve
r applications, where you need all the performance that you can possibly get and where FC-AL's fault tolerance is a major selling point. Superlong cable runs also have some special applications in this arena, such as putting mirrored disks in two separate buildings for additional disaster protection.
There are three reasons why FC-AL is not emerging as a favorite for specialized desktop/consumer applications (video editing, for example). First, a single-user desktop application running on a standard Intel box might not be able to take advantage of FC-AL's speed. (That will change as platforms evolve; many industry experts expect to see FC-AL on desktops by the year 2000.) Second, fault tolerance and disaster protection are not as critical for the desktop/consumer market. Finally, the largest segment of the single-user market, the consumer category, is extremely cost-sensitive, and FC-AL is shaping up to be a bit costly.
FCL: FC-AL + SSA
In September 1996, controller manufacturer A
daptec, IBM's Storage Systems Division, and Seagate announced they would merge FC-AL and SSA into a specification that is now being called Fibre Channel Loop (FCL). (It was initially called Fibre Channel-Enhanced Loop.) With this agreement, IBM effectively admitted that FC-AL had defeated SSA.
But the FCL specification is still in the hand-waving stage. "FCL doesn't really exist," says Joel Warford, business development manager for Adaptec's Mission Critical Subsystems business unit. "It's just sort of a code word for a future serial interconnect that merges the best of Fibre Channel and SSA. Products are years away. Some manufacturers will lead the market in late 1998 and 1999. Others will wait until 2002. It's far enough out that it's difficult to say, but FCL may be a mainstream technology around the year 2000."
FCL is a great deal closer to FC-AL than to SSA. For instance, FCL offers a base speed of 100 MBps -- the same as FC-AL. The FCL proposal was submitted to the ANSI X3T11 committee, the on
e responsible for Fibre Channel. Backward compatibility with FC-AL will be handled through a dual-mode capability, in which the same interface will have the ability to accommodate FC-AL and FCL devices alike. In contrast, SSA devices will not be able to connect directly to an FCL interface. Backward compatibility will have to be handled through combination adapters that have both interfaces on them, such as a gateway.
SSA's probable contributions include
spatial-reuse
and an
arbitrationless
architecture. Spatial reuse allows drives on an SSA loop to talk directly to one another independently of the computer system or the SSA controller. An arbitrationless architecture allows multiple drives to talk to the bus at the same time.
FireWire: Video Va-Voom
FireWire
(IEEE 1394) offers some of FC-AL's benefits, such as hot-pluggability and slim cables. But FireWire is also substantially slower than FC-AL. The current 1394 standard (1394-1995) support
s data transfer rates of 100, 200, and 400 Mbps, or 12.5, 25, and 50 MBps, respectively.
FireWire is inferior to FC-AL in a number of other ways as well. For instance, it currently supports only 63 devices on a single bus and cable lengths among devices up to 14 feet. Neither of those limitations is likely to matter much in a single-user application.
On the other hand, FireWire has characteristics that make it more attractive than FC-AL for certain applications, such as consumer and "prosumer" digital imaging. This is no accident: These are exactly the types of applications for which FireWire was originally designed.
In the last half of 1996, IBM brought out 200-Mbps FireWire transceiver chips, and Adaptec and Skipstone (Austin, TX) brought out 200 Mbps FireWire boards for OEMs. Adaptec should have FireWire PCI boards for end users early this year, for a price of around $350 to $400 -- about the same as the price of a high-end SCSI board. FC-AL boards, in contrast, will be closer to $1000, accordi
ng to Adaptec's Warford. (On the other hand, QLogic displayed a Fibre Channel adapter at Fall Comdex that it expected to sell for about $500.) Adaptec and Skipstone also announced partnerships with digital-video-editing product vendors for products to appear in the first half of this year. Both vendors expect to have 400-Mbps FireWire products late this year.
Compaq, Intel, and Microsoft are all supporting FireWire as a strategic interface. Most observers expect FireWire to appear on Intel motherboards around 1998. "You may see some in 1997," says Allan, "but they have to give disk manufacturers time to switch over to 1394. So, they'll have to leave the current hard disk interfaces in there as well."
Initial FireWire peripherals will include tape decks, digital camcorders, and digital videodisc (DVD) players. No drive manufacturers have yet announced FireWire products.
In spring 1996, a FireWire splinter group formed: the 1394.2 study group. The group's mission is to create a version of 1394 that
can sustain data transfer rates of 100 to 400 MBps, using fiber-optic cable rather than copper wire. Apple, Intel, and Sun have all been promoters of the proposed new specification.
Unfortunately, 1394.2 is incompatible with the current version of FireWire at the physical-interface level. If you have a device (e.g., a digital videocamera) with a 1394-1995 port, you'd need a bridging device to connect it to a microcomputer with a 1394.2 port, for instance.
The companies that have staked their fortunes on 1394-1995 (which is also known as 1394a or 1394.0) are not overjoyed at the emergence of a physically incompatible standard, which would raise prices in a highly cost-sensitive market. They are currently looking into ways of getting 800 Mbps or even 1600 Mbps (that is, equaling or surpassing the 1394.2 proposal) while maintaining physical compatibility. "We see 1394.2 as being more for server/workstation applications, including clustering," says Steve Timm, Microsoft's 1394 evangelist, "not for volume
PC applications. The industry is splintered on that. But 1394.2 doesn't have wide industry support as a motherboard standard. For storage applications, there is a lot of support for making sure that the solution is backward-compatible."
However this battle plays out, the same software will be able to interface to either flavor of 1394 transparently.
SCSI Isn't Dead -- Yet
Of course, something as entrenched as
SCSI doesn't
just dry up and blow away. For a while, SCSI will keep the bulk of its current market -- low- to medium-performance disks. The serial interfaces will divide the spoils as follows: USB gets the mundane things (mice, keyboards, modems, and so on), FC-AL gets disk storage for high-end servers and workstations, and FireWire gets the fun stuff (e.g., digital video cameras, digital videodisc players, and the like).
In the end, FireWire or Fibre Channel (or both) will supplant SCSI. FireWire will get a boost if it comes free on the motherboard -
- especially if it's an 800-Mbps version. From the other side, there's no sign that Fibre Channel prices will come down enough to take any serious market share away from SCSI anytime before the year 2000.
According to Sundi Sundaresh, vice president and general manager for the personal I/O business at Adaptec, "The future of I/O is serial. The question is how long parallel technology will last after the year 2000."
Where to Find
Micropolis
Chatsworth, CA
Phone: (818) 709-3300
Internet:
http://www.microp.com
Superfast Peripherals
