I/O, I/O, the Channel Is Too Slow

I/O, I/O, the Channel Is Too Slow

Dave Tovey

QLogic Corp.

Once a simple choice, selecting the appropriate I/O channel has become a challenge for storage integrators. It is no longer as easy as choosing IDE for PCs and SCSI for servers and workstations.

As applications have increased in complexity and processor power has multiplied, IDE and SCSI have continued to provide timely, backward-compatible enhancements to accommodate demand for increased I/O bandwidth. For example, 16.7 MBps IDE evolved to 33 MBps Ultra ATA, while 5 MBps SCSI evolved to 20 MBps Fast Wide SCSI and to 40 MBps Wide Ultra SCSI.

Today, there`s a whole new list of busses and buzzwords to complement traditional solutions: Fibre Channel, SSA, Firewire, LVD, FC-AL, Aaron, TORN, 1394-1995, 1394.1, 1394a, 1394b, 1394.2, and many more.

Where do they all fit? In some cases, despite the hype, they may not fit at all. Firewire, for example, may not find an early home inside PCs. In other cases, such as Fibre Channel, widespread industry acceptance is a fait accompli.

A key feature of I/O channels is bandwidth: the higher the bandwidth, the more data that can be transferred per second. Traditional I/O channels increased bandwidth by concurrently sharing the data flow among multiple signal paths. Though these "parallel" interfaces (e.g., SCSI and IDE) have served us well for many years, they still have several disadvantages, including high cable and connector costs, relatively short cable lengths, limited device interconnect capabilities and, ultimately, limited bandwidth due to skew effects among the multiple signal paths. Serial interfaces transfer data at very high speeds via a single signal path, thus overcoming these drawbacks.

While widespread acceptance of serial interfaces and the new Ultra2 SCSI parallel interface may appear imminent, integrators should be wary. History shows that new I/O solutions are successful only if there is a compelling reason to change from an existing solution or if there is an emerging application for which there is no current solution. Therefore, new interfaces will always have varying degrees of success.

More Fibre in Your Diet

As a next-generation interface standard, Fibre Channel (FC-AL initially) leads the pack in market acceptance. Supporting up to 126 devices at 100 MBps over 10 km of cable, the Fibre Channel serial interface is a superior solution for high-speed local and remote box-to-box connections and allows greater connectivity in large RAID array or JBOD configurations. In addition, the promise of Fibre Channel`s network capabilities, though not yet a driving force, provides an extra bonus for network storage specialists.

One Fibre Channel hard-drive manufacturer--Seagate--is already in production, and several others are expected to announce production schedules for their Fibre Channel drives by the end of 1997. In addition, the migration to higher levels of integration with embedded transceivers is rapidly reducing the cost of supporting host adapters. And the current high level of I2O activity allows the host system to transparently choose among a variety of SCSI or Fibre Channel adapter solutions while retaining common driver software.

So, all the pieces are in place to enable rapid growth of the Fibre Channel market. Meanwhile, the Fibre Channel Loop (FCL) committee is forging ahead to define a migration path to double and quadruple data rates, to reduce overhead, and to add isochronous capability and bandwidth-enhancing spatial reuse.

Serial Storage Architecture has many of the benefits of a serial interface, but it has been relegated to a solution primarily for IBM`s installed base. However, its better features will likely be incorporated into next-generation Fibre Channel specifications, probably late in the decade. Meanwhile, other interface proposals such as "Aaron" and "TORN" have been absorbed by the Fibre Channel working committee, with their best features retained for use in future FCL specifications.

Extending SCSI

Ultra2 SCSI, enabled by low-voltage differential (LVD) technology, is the next SCSI evolution, doubling the transfer rate to 80 MBps. To meet the ever-increasing bandwidth demands of workstations, servers, and disk drives, Ultra2 SCSI will begin replacing Ultra SCSI in late 1998. However, some proponents are pushing to use Ultra2 SCSI in earlier applications that clearly favor a Fibre Channel solution.

While Ultra2 offers considerable bandwidth gains, it is no match for Fibre Channel`s connectivity, low cable and connector costs, long cable length, and bandwidth migration path. Also, Ultra2 has yet to be approved by the ANSI committee, while Fibre Channel has three-and-a-half years of specification stability and implementation. The message is simple: Ultra2 SCSI is the Ultra SCSI successor, not the Fibre Channel alternative.

Enough Fire in the Wire

On the desktop, IEEE 1394 ("Firewire") proponents are rallying to replace the venerable IDE interface. But IDE`s data rates (currently, 33 MBps for ATA Ultra and, in the future, 66 MBps for ATA Ultra2) can support at least the next five generations of PC disk drives. Moreover, these interfaces are backward compatible with existing hardware and software, are low cost, and have a proven track record.

Therefore, without offering a compelling reason to change, 1394 faces an uphill battle in trying to displace IDE inside the PC box and thus will unlikely penetrate the market before 2000.

Future PCs, however, may take the form of "sealed boxes" with external 1394-based "Device Bays" for peripheral expansion and thereby would provide an improved add-on solution that could generate a respectable, albeit still questionable, market for 1394.

In addition, significant industry debate among proponents of the partially compatible implementations (1394-1995, 1394a, 1394b, 1394.2, etc.) will further delay widespread acceptance.

Nonetheless, 1394 has already found a home in the consumer digital video interconnect market. With isochronous capability, 1394 can support multiple concurrent, guaranteed on-time delivery video channels on a single small cable connecting a digital camcorder, DVD, set-top box, HDTV, digital-still camera, and other devices.

Just how quickly this market will grow remains to be seen. However, current predictions suggest Japan will be the initial market, followed by Europe and North America around the year 2000.

IEEE 1394-1995 Firewire offers speeds of up to 400 Mbps, with a proposed migration path to 800, 1600 and 3200 Mbps via 1394b. However, cable and protocol compatibility issues have not yet been fully resolved. Throw in 1394a, which would improve overhead, and you have an uncertain situation that predicates market delays. And finally, there`s 1394.2, which promises Fibre Channel-like bandwidth at a PC market price but is not compatible with other 1394 solutions.

The bottom line: Despite the hype, don`t expect 1394 to seriously penetrate the market in the next two to three years.

Easing the Transition

As the data storage industry migrates to new I/O solutions, the transition will be facilitated by widespread acceptance of I2O. Until now only supported by unique solutions from one or two focused suppliers, the I2O concept of an I/O-platform-transparent interface between the O/S and I/O processor will become widely available from multiple suppliers and will significantly reduce the cost and risks of upgrading I/O performance and connectivity.

And what about beyond 2000? The PC will most likely evolve to a sealed box with two serial connectors: USB for accessories (keyboard, mouse, etc.) and 1394 or Fibre Channel (or even some other low-cost serial interface) for all other peripherals. For workstations and servers, Fibre Channel will dominate, but parallel SCSI will still be an effective solution. Ultra3 parallel SCSI, with a 160 MBps bandwidth, will also be found extensively inside RAID boxes and other drive subsystems. Large RAID systems, box-to-box connections, and video arrays will predominantly use Fibre Channel.

So, will I/O channel decisions get any easier? Probably not, since there`s sure to be a whole new set of contenders--all of which will promote faster transfer rates, lower costs, and ready availability. The strong will survive, but as usual, it will take two years longer than expected for any new interface to penetrate the market.

And so the tune will continue ... I/O, I/O, the channel is too slow.

Dave Tovey is vice president and general manager, peripheral products, at QLogic Corp. in Costa Mesa, CA.

This article was originally published on October 01, 1997