Design Feature: October 12, 1995

Advances in peripheral interfaces—some gradual, some radical—are now giving designers a wealth of choices for high-speed, feature-laden I/O. And making good choices based on the features of each interface and your system requirements is paramount to maximizing system performance and device compatibility. A thorough evaluation, however, must also consider the maturity of each technology to ensure that the advanced features an interface offers are also affordable and available.

Over the last half-dozen years, interface choice—at least for primary disk storage—has been relatively simple. PCs based on Intel x86 processors have predominantly used IDE (integrated-device-electronics) interfaces for disk drives. In rare cases, x86-based systems have turned to SCSI (Small Computer System Interface). Generally, these instances have been limited to applications in network file servers or in systems that use an inherently multitasking operating system, such as Unix. These applications exploit the capabilities of SCSI disk drives to handle queued commands and the generally faster data-transfer rates afforded by SCSI drives.

Outside the x86-based PC world, SCSI has dominated as the disk interface of choice for some time. Apple Macintosh systems have always used SCSI as a disk interface, as have workstations, VMEbus-based embedded systems, and even enterprisewide servers and transaction-processing systems.

The old rules no longer apply in any of these applications, however. SCSI originally struggled to make an impact in the PC world, not because of cost or performance, but because of difficulties in getting the interface to work in an MS-DOS world that included no inherent support for SCSI. With Windows 95, the support is finally here. Moreover, the multithreaded, multitasking Windows 95 operating system can truly leverage the capabilities of SCSI.

Further clouding the issue, IDE and SCSI sport recent enhancements. Furthermore, several new interfaces including SSA (serial storage architecture), FC-AL (Fibre Channel arbitrated loop), P1394 FireWire, and USB (Universal Serial Bus) are in the early stages of availability and will be candidates for usage in 1996.

Data rates soar

You can attribute the moving and shaking in the peripheral-interface arena to two factors. First, manufacturers are forcing disk drives to deliver data at faster rates to keep pace with faster µPs and memory. The fastest disk drives are starting to saturate the capabilities of the interfaces in applications from PCs to dedicated-storage subsystems, such as disk arrays.

Second, the sheer range of peripherals available for computers has created a need for more diverse interfaces. Examples include CD-ROM drives; tape drives; and rewritable, rotating storage devices, such as magneto-optical drives; removable hard-disk drives; and high-capacity floppy-disk drives. Besides these secondary storage devices, other peripherals—including scanners; video cameras; and the ever-present pointing device, keyboard, and printer—also require connections.

It's clear that providing a dedicated interface to every peripheral is impractical. On the other hand, a single interface can't likely serve all peripherals, because you cannot allow slower secondary peripherals to hamper the performance of the primary disk subsystem and other high-speed devices (see box, "Utility buses"). Whether you are working with a PC architecture, a workstation-class design, or even a disk array, your first job is to discern the best way to interface the fastest peripherals in a system. You can then decide whether that primary interface can also serve other peripherals and what other secondary interfaces your design will require.

Enhancements to IDE

A consideration of disk interfaces should start with IDE or, more correctly, enhanced IDE (EIDE) because approximately 70% of all drives built today use the EIDE interface. Due to this volume and a low-cost design focus, EIDE drives cost less than drives with other interfaces in capacities of 1.2 Gbytes and below.

IDE, also known as "ATA" for ATbus attachment, became dominant in the PC market because it directly replaced the ST-506 controller used in the IBM PC/AT. The IDE scheme simply moved the disk-control function from the ST-506 ISA bus card directly onto the drive. The integrated controller allowed for features such as cache on the drive and boosted capacities through techniques such as zone bit recording, largely innovations from SCSI drives that had long integrated control functions on the drive.

IDE, however, had some limitations in that the interface could support only two drives and data rates were maxed out at 2 to 3 Mbytes/sec, even when coupled to the system via a local bus. The EIDE movement addresses shortcomings in the original IDE spec in four areas: disk capacity, transfer rate, number of devices supported, and support for peripherals other than disk drives.

EIDE is a de facto name, not a formal standard, that the industry regularly applies to advancements in the above areas. EIDE-labeled boards and ICs almost always support enhancements in these areas. The Small Form Factor (SFF) Committee, a storage-industry standards committee, publishes pertinent specs.

Solving capacity limits

IDE had no severe limits in disk capacity. The original spec could handle drive capacities greater than 100 Gbytes. The combination of the IDE spec and the typical system BIOS limitations created the artificial 528-Mbyte limitation in the PC architecture (Table 1). New enhanced-BIOS implementations have solved these capacity constraints by translating physical cylinder, head, and sector addresses to logical-block addresses (LBAs), much like those that SCSI drives use. EIDE drives include task file registers that store LBA translation tables and ensure support for drives up to 8.4 Gbytes in capacity.

Boosting performance in the EIDE interface was a bigger issue for system manufacturers attempting to maximize performance. When IDE was defined, disk drives could provide continuous data rates of less than 1 Mbyte/sec, whereas today's top-end EIDE drives leverage significantly higher area density and rotational speeds to offer transfer rates in excess of 6 Mbytes/sec.

EIDE data rates

Boosting performance in the EIDE interface is a multifaceted effort. EIDE allows several forms of DMA transfers, although the IDE interface relied strictly on programmed I/O and the host processor for data transfers. EIDE also includes a burst-transfer mode, called Mode 3 PIO that uses programmed I/O commands and boosts rates to 11 Mbytes/sec. The fastest EIDE transfers come from a new DMA bus-master transfer mode that relies on a PCI bus and offers rates as fast as 16.6 Mbytes/sec. Drives and interfaces that support this DMA-master mode should appear over the next six months.

EIDE efforts to support more total devices was more an effort to coordinate support with BIOS authors and chip vendors than to fundamentally change the interface. The PC/AT architecture has always supported two disk controllers, each of which could control two drives. Some BIOS implementations, however, lacked support for the second controller.

An enhanced BIOS can now support two EIDE interfaces. Moreover, most EIDE-interface-chip vendors, including core-logic chip-set vendors that directly support EIDE, have added the hardware necessary for a second interface. In turn, motherboards and EIDE add-in cards now typically carry two EIDE connectors. The move to support four devices adds about $1 to the price of a system.

The final enhancement to IDE is support for additional device types, including CD-ROM drives, tape drives, and even floppy-disk drives. EIDE supports these devices by defining new controller-operation codes and The AT Attachment Packet Interface (ATAPI), a logical-packet interface that can be layered onto other EIDE capabilities. The SCSI-like ATAPI command set doesn't interfere with normal disk activity, although a slow device, such as a CD-ROM, could hog the bus and slow disk activity. For performance reasons, most designers want to reserve the primary EIDE channel for hard-disk drives and the secondary channel for slower devices.

All the EIDE activity has spurred interest outside the Intel-based PC world. For example, Apple is now shipping Macintosh systems that use a PCI bus to host an EIDE interface for disk storage. The systems still include a SCSI port to link other peripherals, but the allure of lower cost hard drives instigated the switch to IDE for the primary disk interface. Workstation vendors, such as Digital Equipment Corp, and embedded-system vendors working on VMEbus platforms have also begun to use EIDE in some cases. Even some disk-array vendors are searching for ways to use the low-cost EIDE drives.

Windows 95 opens door for SCSI

With all this momentum behind the EIDE interface, you might think that SCSI still has no place in Intel-based PCs. The new Windows 95 operating system, however, along with growing popularity for other multitasking systems, such as Windows NT, make SCSI a better choice than EIDE, based strictly on performance. Moreover, Windows 95's plug-and-play capabilities have simplified the task of adding SCSI to an x86-based PC. And a new SFF committee, SCSI Configured Auto-Magically (SCAM) has Microsoft's backing and promises to further simplify matters by defining a scheme for SCSI devices to probe the bus, set their own addresses, and determine whether their interface should terminate the bus.

SCSI offers a number of performance advantages over EIDE. First, a single 8-bit SCSI interface can connect as many as eight devices: the host plus seven peripherals and a 16-bit interface can connect 16 devices. Second, readily available 8-bit Fast SCSI drives transfer data at 10 Mbytes/sec, and 16-bit Wide SCSI drives extend that speed to 20 Mbytes/sec. Even faster SCSI implementations are coming soon.

Of even greater importance to system performance, however, is a host system's ability to simultaneously issue multiple commands to one or more SCSI devices. In other words, a multithreaded, multitasking operating system allows tasks to have commands simultaneously active in the disk subsystem. In contrast, MS-DOS, Windows 3.1, and the Macintosh operating system issue a command to the disk subsystem and must wait for the command to complete before issuing another. The ability to overlap commands and queue the commands to one or more devices can significantly boost performance in environments such as Windows 95 and NT.

SCSI could be poised for rapid growth in the PC market if it can overcome the cost issue. Realistically, a SCSI-disk subsystem should not cost significantly more than an EIDE subsystem. The SCSI interface on the host side should add less than $50 to system cost and replace the need for additional interfaces.

EIDE/SCSI price trade-offs

The choice of SCSI or IDE for disk drives comes down to one IC vs another. The EIDE controller should be less expensive due to a less complicated controller IC. The difference, however, should be ones or tens of dollars rather than the $100 or more difference in street prices of 1-Gbyte drives. Until two or three years ago, the price difference between SCSI and IDE devices was $10 to $20. IDE and SCSI drivers were identical, except for their interfaces. Over the past few years, however, disk manufacturers have separated IDE and SCSI product families in terms of performance in an attempt to minimize IDE drive costs.

For example, a manufacturer might offer a 540-Mbyte EIDE drive with a 12-msec average seek time and a 540-Mbyte SCSI with a less-than-10-msec average seek time. Likewise, the SCSI offering may have a 10 to 20% faster spin rate and, therefore, a faster data-transfer rate. You can view the greater performance of the SCSI drives as a bonus. The price difference is a tremendous obstacle, however, in the price-sensitive PC market.

The price difference does appear to be shrinking. Perhaps the drive vendors have recognized the potential of SCSI in Windows 95 systems or perhaps the short-term decline in the price of SCSI drives is a simple anomaly in the price curve. Quantum, for example, is trying to force together the price and performance of EIDE and SCSI drives. On the other hand, some vendors, such as Maxtor, simply quit manufacturing SCSI drives to concentrate on EIDE products.

SCSI extensions

SCSI does offer one thing that EIDE can't: even greater headroom for further extension. SCSI will continue to dominate for the foreseeable future outside the realm of the PC. Part of the reason for that dominance is continued enhancements to the SCSI spec. SCSI-3, the latest official release of the SCSI specification, marks a significant departure from the previous structure of the interface. SCSI-3 comprises several specs that handle specific layers. For example, one spec deals explicitly with the logical-command layer. You can now use the command layer with a variety of physical layers (Table 2).

The new structure allows the ANSI and SFF committees to change or enhance one area without affecting other areas. This increased freedom should mean that enhancements get formal approval faster.

Fast-20 doubles data rate

The 8- and 16-bit implementations with support for fast transfers currently predominate. Over the past six months, however, the Fast-20, or UltraSCSI, enhancement has surfaced in a handful of products that can support 20-Mbyte/sec, 8-bit rates and 40-Mbyte/sec, 16-bit rates. Fast-20 doubles the data rate over regular Fast SCSI by doubling the synchronous clock speed.

Fast-20 originated in Digital Equipment's disk-drive division before Quantum purchased that division last year. Not surprisingly, Quantum this summer became the first manufacturer to ship a Fast-20 disk drive. Symbios Logic (formerly NCR Microelectronics) and Western Digital now offer both IC and host adapters for Fast-20 systems, and Adaptec and BusLogic have introduced host adapters.

Yet another enhancement, Fast-40, is in the pipeline. As you might guess, Fast-40 promises to double data rates once again. Fast-40 uses a differential signaling scheme that is necessary to double the clock speed again.

Fast-40 differential

SCSI dedicates a ground pin for each data pin on the interface connector, so Fast-40 proposes using that ground pin to enable differential signaling. SCSI also includes an option for differential transmissions, but the option supports long cable runs (25m compared with 6m for single-ended implementations). The Fast-40 proposal, however, will use lower power, lower cost differential transceivers than existing differential SCSI. IC vendors expect to support Fast-20 and Fast-40 with the same IC.

Fast-40 adds another option to storage subsystems. Although Fast-20 systems ideally could still support a full complement of devices, the faster clock speed and capacitive loading realistically limit a Fast-20 interface to three to six devices. A Fast-40 differential implementation operating at Fast-20 speeds could handle 16 devices on a cable.

Whether Fast-40 is the top end for parallel SCSI or just a plateau, the computer and storage industries have been exploring new physical interfaces that use serial physical layers. Interestingly, any such interfaces will likely still use the SCSI-3 command layer. You've probably heard of the two leading serial contenders—SSA and FC-AL—because overzealous marketers have been touting the virtues of each at every opportunity.

Dal Allan, SFF chairman and president of ENDL Consulting, warns, "Watch out for the 'true lies'" from the serial proponents, especially when they are taking shots at each other. For example, FC-AL proponents claim that their interface offers 200-Mbyte/sec performance compared to 20 Mbytes/ sec for SSA. SSA proponents, on the other hand, tally the comparison at 100 Mbytes/sec for FC-AL and 80 Mbytes/sec for SSA. The truth lies somewhere in the middle; moreover, a straight comparison on maximum bandwidth tells little of the story.

Some SSA proponents are also taking on parallel SCSI, touting higher performance, cheaper cable and connectors, and, ultimately, cheaper serial transceiver ICs. Any such price advantages, however, are at least five years in the future. Parallel SCSI is low on the price curve, and its connectors, ICs, and even backplanes that eliminate cable-transmission problems are affordable. For the next few years, the serial interfaces will have to earn design wins through better performance. The use of a serial interface as a performance upgrade to a parallel interface is at first hard for engineers to accept. They've learned for years that parallel interfaces are generally faster.

Serial vs parallel

Serial techniques make sense for two reasons. First, an affordable serial interface lets you use more capable transceivers to handle the higher speeds and simultaneously greater cable lengths. Second, serial interfaces eliminate the difficulties in accurately clocking and capturing data over eight, 16, or even 32 signal paths in cables with differing latencies and transmission-line effects impacting each signal path. Although SSA and FC-AL both use serial transmissions, the two differ widely.

FC-AL's strength lies in its ability to transfer data from one point to another at 100 Mbytes/sec. SSA, conversely, can provide only 20-Mbyte/sec rates between any two points but supports multiple simultaneous 20-Mbyte/sec conversations around a ring (Fig 1). The scheme is based on inherently full-duplex connections between two points that can handle 20-Mbyte/sec rates in each direction. You can use SSA in a string configuration, but the loop, as in Fig 1, will likely become the more popular implementation.

A loop topology requires that each device have dual SSA interfaces, eliminates any single point of failure on the loop, and lets you add devices to the loop without powering down the system. Moreover, an initiator on a healthy dual loop can simultaneously use both full-duplex ports making the theoretical loop bandwidth 80 Mbytes/sec. Even this number smacks of "specsmanship," but proponents claim that the loop can deliver even more than 80 Mbytes/sec through "spatial reuse." SSA, for example, would theoretically allow the initiator in Fig 1 to maintain full-speed, full-duplex communications with its adjacent nodes, while the targets on the other side of the loop communicated among themselves.

Fibre Channel also comes in many flavors (Fig 2). The switched fabric was originally conceived as the most likely implementation and still holds promise for applications such as multiprocessing. Fibre Channel lets you use a variety of high-level protocols on the physical connection. For example, companies in the data-communications business are layering Transmission Control Protocol/Internet Protocol layered onto a Fibre-Channel fabric. The loop-based FC-AL Direct Attach Disk Profile, which ANSI Committee X3T11 specified, is the topology of choice for the data-storage industry.

Designers can use single- or dual-loop FC-AL implementations. The dual-loop implementation protects against a single point of failure and offers a theoretical 200-Mbyte/sec bandwidth. Although diagrams of SSA and FC-AL loops are similar, the differences are substantial. In FC-AL, one member of a loop arbitrates for control and then sends data around the loop to the appropriate target. FC-AL allows only one data transfer per loop at a time.

SSA, on the other hand, uses no arbitration. Each SSA connection is a true point-to-point link. Device A transmits a data packet to Device B. A three-way router in Device B examines the packet and decides whether to accept or retransmit the data.

SSA or FC-AL?

It's currently impossible to say whether SSA or FC-AL is better for your long-term plans. You should select the one that ultimately garners the most support and, therefore, the widest choice of products at the lowest price. It could be three years before one technology will gain the needed support. Both have impressive backers. IBM has championed SSA and has enlisted a number of system, peripheral, host-adapter, IC, and even connector manufacturers. A formal marketing/standards organization, the SSA Industry Association carries the SSA flag.

Meanwhile, three leaders in high-end disk drives—Hewlett-Packard, Quantum, and Seagate—along with board and IC vendors back FC-AL. Most disk vendors are taking an either-or stance. Only Conner Peripherals among the leading companies has pledged to support both standards. IC and board vendors, however, will in many cases support both for the foreseeable future. Today, you can buy SSA drives from IBM and Conner. Conner, Hewlett-Packard, Quantum, and Seagate are starting to ship samples of FC-AL drives.

Basing your choice on the merits of the interfaces can be as simple as examining your application. SSA will likely be less expensive for some time, because manufacturers are building SSA ICs using a 0.5-µm CMOS process. FC-AL ICs, on the other hand, require either a 0.35-µm CMOS process or the use of GaAs ICs. This difference will also mean that SSA ICs are more readily available in the near term because few 0.35-µm fab lines are available. FC-AL is probably the best choice if you need to quickly move large, rich data streams for applications such as high-end graphics or video. SSA could be better for handling many small I/O transactions, although SSA can become saturated as you add devices. FC-AL can effectively support more devices due to its raw bandwidth.

You should also consider the system-level complexity of using one of these interfaces before rejecting parallel SCSI. Even though you can map SCSI commands to run on these interfaces, you may have to make system and software changes to realize optimal performance. For example, you would have to change current operating systems and drivers to take full advantage of the dual SSA ports and the concept of spatial reuse. Likewise, you would have to make fundamental changes to the software drivers to take advantage of a dual-loop FC-AL implementation and be able to revert to using one loop if a device failed.

Take another look at your entire system before using any of these enhanced interfaces—EIDE with bus-master DMA in a PC, Fast-40 in a workstation, or FC-AL in a video server. It may be time to revamp your I/O subsystem, because both disk drives and interfaces may now be an order of magnitude faster than the speed for which the operating-system driver and host interface were designed to handle.

You can reach Technical Editor Maury Wright at (619) 748-6785, fax (619) 679-1861.