May 8, 1997

SCSI evolves to meet changing, challenging demands of small computers

Erich Otto, Symbios Logic Europe GmbH

The SCSI standard committee refuses to stand by and watch advances in the small-computer market pass it by. SCSI advances have met the high-speed challenge head-on.

The SCSI (Small Computer Systems Interface) standard, now in its third official iteration, has been an important part of the growing small-computer market over the last decade and a half. The standard has provided a flexible, high-speed interface for connecting a variety of peripherals to several types of hosts. The flexibility of the SCSI bus is characteristic not only of the interface but also of the standard itself. The SCSI standard has been able to remain a viable interface from the early 1980s with the SCSI-1 specification to the SCSI-3 specification today, which is completing development. The standard has remained competitive by evolving from a slower asynchronous-only bus to the upcoming high-speed, parallel Ultra2 SCSI and serial SCSI, such as Fibre Channel, Serial Storage Architecture (SSA), and 1394.

The SCSI standard was developed because of the need for a high-performance, flexible peripheral bus that freed the host processor from handling much of the data distribution. In 1981, NCR Microelectronics (now Symbios Logic) joined Shugart Associates to develop the Shugart Associates Systems Interface (SASI) specification. In 1982, when the X3T9 Standards Committee adopted this standard, it was renamed SCSI. After many years of work and many more revisions, ANSI accepted the SCSI-1 specification in 1986. The evolution of SCSI didn't end there. Proposals for the SCSI-2 were already in the works.

The legacy interface

Most peripherals and host adapters currently in use are from the SCSI-2 era; most of the SCSI-1 devices have died out because of inadequate capacity or performance. SCSI-2 provides several advantages over the older devices with fast transfers, wide bus implementations, differential signaling, multitasking, and multi-initiator systems.

Although the SCSI-1 specification did support synchronous data transfers, it failed to take full advantage of its potential, which is still expanding. SCSI-1 synchronous transfers were comparable to the existing slow asynchronous transfers. SCSI-2 ex-pands this performance to 10 megatransfers/sec (10 Mbytes/sec on a standard 8-bit bus) and is the first to take true advantage of the synchronous signaling. These higher transfer rates are known as Fast SCSI. With the expanded use of higher speed synchronous transfers, a special type of SCSI driver, which uses active negation, was recommended. These drivers help provide clean signaling and faster transitions during Fast SCSI transfers.

Wide SCSI-bus implementations offer better performance using existing technology as long as price and size are not issues in an application. The SCSI-2 standard provides for 8-, 16-, and 32-bit data-bus widths, but only the 8- and 16-bit versions are widely accepted. When applications started to move to the 16-bit version (commonly referred to as Wide SCSI), the data rates effectively doubled to 20 Mbytes/sec using only the existing Fast SCSI technology. In most new developments, Wide SCSI is becoming the interface of choice as the connector form factor and the prices come closer to those of traditional 8-bit SCSI.

Differential signaling also became an option in the SCSI-2 standard, which solved many electrical issues with the SCSI bus but added significant cost. Differential signaling allows the use of two conductors to represent one logical level, which allows cleaner signals compared with the traditional single-ended environment. Another advantage is longer cables. You can extend differential buses to 25m, compared to the 3m restriction of single ended. Price is the major obstacle for wide acceptance of differential devices. Because differential transceivers cannot be incorporated into a standard SCSI controller chip, you need additional components, which raises the application cost. Therefore, only very high-end applications have chosen to implement differential.

SCSI-2 also increased its performance and flexibility through other features, such as multitasking, a new message set called Tagged Command Queuing, and multi-initiator systems, which allows peripheral sharing. These and other new features allowed SCSI to not only survive its second generation and its use into the 1990s, but also to expand its acceptance from the industrial market to the PC and workstation markets.

PnP: making SCSI easier to use

As SCSI now moves into its third generation (the SCSI-3 specification is still under development), a number of additional improvements will allow it to remain competitive. One of the first is the SCSI Plug 'n Play (PnP) profile, which follows the trend to allow average computer users to easily and effectively install host adapters and peripherals.

The SCSI PnP specification covers many of the aspects of SCSI that tend to confuse or make a SCSI device harder to implement; it standardizes the various physical components, such as cables and connectors. The SCSI-1 and SCSI-2 specifications provided for a flexible interface, which includes different connector types. Unfortunately, this flexibility sometimes backfires and requires the user to have several cables available to ensure that one fits. The new spec standardizes the 50-position (68-position for wide devices), high-density, shielded connector for external connections. Other connections are still valid within the SCSI specification but cannot claim PnP compliance.

The termination of the SCSI bus is always an important consideration in the parallel-SCSI environment, but now with the faster standards beyond the traditional Fast SCSI, the consideration becomes doubly significant. To help, the PnP specification introduces a standard cabling environment and an autotermination requirement for host adapters. Peripherals external to the PC enclosure must be daisy-chained with a terminator at the end of the chain. Also, for host adapters that provide a connection to both internal and external peripherals, an autotermination scheme is required to detect whether the host adapter is in the middle or at the end of the bus. These considerations simplify the computer-knowledge requirements of the end user.

Finally, for new users trying to configure a SCSI system, the questions regarding SCSI ID numbers always come up. What are they supposed to be set to? Am I supposed to match each of the numbers? A part of the SCSI-3 standard called SCAM, or SCSI Configured AutoMatically, addresses and in most cases eliminates these questions. It assigns each device a SCSI ID number based on a unique device ID number.

Two levels are associated with SCAM. SCAM Level-1, the most commonly implemented, allows the automatic assignment of SCSI ID numbers to peripherals in a single-initiator system. SCAM Level-2 goes beyond to support additional features, such as multi-initiator systems and hot-plugging.

When all of these aspects of the SCSI PnP profile are implemented, the configuration of a SCSI bus is much easier. The SCSI PnP does not quite make the SCSI bus foolproof, but it's a step in the right direction. (Some say that PnP is the next step in the race between the computer industry to make foolproof systems and the world to make better fools.)

Ultra SCSI: doubling the speed

Although Fast SCSI-2 devices dominates the installed base of SCSI systems, the new Ultra SCSI dominates new development. Ultra SCSI, previously known as Fast-20, provides an easy migration path for users to nearly double the performance of their systems using a similar technology, allowing data-transfer rates to 40 Mbytes/sec on a Wide SCSI bus.

Because of a similar technology in Ultra SCSI, backward compatibility is easy to achieve. Although many of the electrical and timing specifications are tighter in Ultra SCSI, Ultra and Fast SCSI still remain essentially the same. From a software and protocol standpoint, the only difference is in the negotiation for the higher transfer speed. Therefore, backward compatibility isn't really an issue in these systems. For example, a system that uses a new Ultra hard disk and host adapter would be able to use an older Fast SCSI CD-ROM and scanner and still get the maximum performance from each.

Ultra SCSI essentially doubles the synchronous-transfer rates during the data phases but leaves all of the other bus phases, such as command and message, to the pre-existing values. For each I/O transfer, the amount of overhead remains the same between Fast and Ultra transfers. Therefore, systems that use large block transfers benefit more from the Ultra transfers than from others (Figure 1). Many times, these block sizes depend on the operating system and applications. Operating systems such as Windows NT and some types of Unix benefit more from Ultra SCSI than, say, DOS and Windows 95 do.

Some of the disadvantages of Ultra SCSI come from tight electrical specifications. Users, for instance, must consider the requirement for a higher quality (and therefore, more expensive) cable; in single-ended systems, you need shorter cable and fewer devices (Table 1). The differential interface remains almost the same among the interfaces because of its improved signaling characteristics.

Ultra2 SCSI: doubled again

Just around the corner is the release of the latest of the parallel-SCSI evolutionary steps, Ultra2 SCSI, as defined by the SCSI-3 Parallel Interconnect 2 standard. Ultra2 SCSI follows closely in the path of Ultra SCSI by taking its transfer capability and doubling it again, which allows for data-transfer rates of 80 Mbytes/sec max on a Wide SCSI bus. But, because the existing single-ended and differential interfaces cannot support these transfer rates, a new interface standard is required.

The SCSI committee adopted low-voltage differential (LVD) for just this purpose. The differential interface has long been known to have better electrical characteristics than do single ended, but the requirement of external transceivers usually makes differential unpopular. LVD improves upon what is now called "high-power differential" (HVD) by requiring lower voltage swings (60-mV differential receiver thresholds) and a 1.25V common-mode biasing. These smaller voltage requirements allow the transceivers to be incorporated into the SCSI controller still using the low-cost CMOS technology, providing a differential solution that only slightly increases cost over single ended.

LVD signaling provides other electrical advantages that make configuring the SCSI bus more flexible. An LVD bus can include the full 16 devices on a Wide SCSI bus and still achieve cable lengths of 12m--a significant advantage over the 1.5m allowed in fully loaded, single-ended Ultra SCSI systems.

Besides cost, another advantage of incorporating the LVD drivers into the controller's silicon is that LVD allows some backward compatibility to single-ended systems. This feature is commonly known as Universal LVD. The hardware detects legacy single-ended devices through the Diffsens signal and automatically switches the drivers to a single-ended configuration. Similarly, LVD is compatible with HVD, but external transceivers again become necessary. The only disadvantage to this backward compatibility is that systems that revert to single ended or HVD are limited to the Ultra SCSI transfer rates.

Serial SCSI: Fibre Channel, SSA, and 1394

Besides the rather natural progression taken by the parallel SCSI interface, a few other serial interfaces have partially evolved from SCSI. Fibre Channel, the SSA, and 1394 have SCSI as a common ancestor. In general, these standards try to apply some of the best features from the parallel interface to a serial one.

The serial SCSI architectures have some advantages over the parallel interfaces, such as longer cable, small cables; and connectors; and, in some instances, higher data-transfer rates. For example, Fibre Channel is currently being designed at 1-Gbps transfer speeds with cable lengths to 10 km. Unfortunately, cost and device availability are holding these interfaces back. Serial interfaces, especially Fibre Channel, will probably replace parallel SCSI, but that scenario is a few years off.

The SCSI committee maintained compatibility with the serial architectures by dividing the SCSI-3 specification into several layers. The two most general classifications are the logical and the physical layers. The physical layer is specific to the parallel interface, and the logical layer can be adapted to almost any interface. A major portion of this logical layer is in the common command set, which provides a common interface to many devices. This flexibility of SCSI has long been one of its most attractive features. Using a common logical layer between the interfaces also aids software and firmware migration between platforms.

Through both the parallel and serial SCSI implementations, additional performance improvements are already being discussed and tested by many members of the SCSI community. One such improvement from the parallel side would implement transfer rates that are double and quadruple the speeds available from Ultra2. The LVD technology suits transfer rates faster than 100 Mbytes/sec with only minor improvements over its current implementation.

Fibre Channel, SSA, and 1394 are moving at a similar development pace. Even before the technology has been widely accepted, developers are looking at higher speeds, support for other protocols, and additional classes of service. The serial interfaces will need to develop quickly to eventually surpass their parallel cousins.

Regardless of the interface, the SCSI industry has some challenges ahead. The computer industry continues to progress at a fantastic rate, and it's SCSI's duty to ensure that the peripheral interface is not the limiting factor.

References

1. Relf, Stanley E, Understanding the Small Computer System Interface. Prentice Hall, Englewood Cliffs, NJ, 1990.

2. SCSI Trade Association World Wide Web: www.scsita.com

3. X3T10 Technical Committee World Wide Web: www.symbios.com/x3t10.

| EDN Access | Feedback | Table of Contents |