If you worry about losing data through hard-disk failures, RAID can put your mind at ease. RAID stands for (take your choice) redundant array of inexpensive/independent/-individual disks/drives. RAID techniques can also accelerate disk-drive data-transfer rates for such speed-intensive applications as video processing.

As Table 1 on pg 86 shows, a range of SCSI-interface RAID products is available for PCs and workstations. The products range from inexpensive controller boards and software packages for those who wish to configure their own RAID systems, to complete self-contained drive subsystems.

This article discusses products at the "low end" of the RAID spectrum, mainly sub-$15,000 systems that are likely to be of interest to designers. High-end systems (note the Core Array 2000 in Table 1), costing many tens of thousands of dollars, are more in the domain of MIS managers and network administrators.

RAID should be of interest to designers in large CAD systems. Joel Hagberg, RAID-product manager for Micropolis, cites the example of engineers sharing files over a network, a common scenario in many design facilities. RAID techniques are appropriate for such systems, both for faster data transfer and for fault tolerance. Although you can use tape-backup systems to provide fault tolerance, they have not kept pace with the advances in drive technology, according to Hagberg. Backing up with tape is agonizingly slow.

Though this article (and Table 1) deals with rotating-disk RAID systems, one emerging solid-state system is also worthy of mention. A 320-Mbyte flash-storage system, using RAID levels 3 and 5, is available from Raymond Engineering Inc. Designed for rugged environments (such as 1500g shock), the system is about the size of a large coffee mug. As usual, however, solid-state mass storage cannot compete with rotating media: The system, available in military versions now, costs $30,000 to $50,000.

What RAID level is best for you?

RAID accelerates data transfer through "striping"--distributing segments of the data across multiple drives. Two RAID techniques provide redundancy in systems requiring fault tolerance: "mirroring" and exclusive-OR parity architectures.

Fig 1 provides a simplified block diagram of the RAID levels in popular usage. RAID 0 (Fig 1a) uses pure striping to achieve high transfer rates. The stripes usually comprise large segments of data. For single-user applications that demand access to long sequential records, however, the stripes can be sector-sized (if the spindles are synchronized).

RAID 0 is the speediest and most disk-efficient of the levels. Its disadvantage is the total lack of redundancy--if one disk fails, the system fails. RAID 1 (Fig 1b), on the other hand, is the least disk-efficient but provides total storage redundancy. In RAID 1, identical data is mirrored to two disks. The disadvantages of RAID 1 are the expense (double the number of disks), and the fact that there's no performance improvement over a single-disk system.

Fig 1 does not illustrate RAID 2, which has fallen out of favor and become all but extinct. This level is identical to RAID 0--data is striped across multiple disks--except for the addition of drives that store error-correction code (ECC) information. Most modern disk drives come with embedded ECC, so RAID 2 serves no useful function.

RAID 3 (Fig 1c) uses striping and provides redundancy via a disk dedicated to storing parity information. In the case of a drive failure, the system recovers data by generating the exclusive OR of the data on the remaining drives. RAID 4 is structurally similar to RAID 3 but uses larger stripes, so that reads can originate from any drive (except the parity drive), thus allowing overlapping of read operations. Another difference between RAID 3 and 4: A RAID 3 array executes one I/O request at a time, involving all member disks equally, and a RAID 4 array can process multiple requests simultaneously, as long as the specified data maps to different member disks (Ref 1).

RAID 3 requires synchronized spindle drives to avoid performance degradation with short records. It offers good throughput rates for large data transfers but relatively poor media efficiency for small data blocks. RAID 3 is appropriate for imaging applications, for example.

A RAID 5 system (Fig 1d) spreads both the data and the parity information across all the drives. Unlike with RAID 3 or 4, each drive takes turns storing parity for a different series of stripes. Because there is no dedicated parity drive, it's possible to overlap read and write operations. The overhead in RAID 5 is the time it takes during write operations to regenerate the parity information. The result is a degradation of 40 to 66% in write speed, as compared with a RAID 1 (mirroring) array. RAID 5 is thus an appropriate choice in multiuser applications that require few write operations.

RAID 5 finds use in many more systems than does RAID 4. The fact that its parity chunks are distributed across multiple array members, rather than being concentrated on a dedicated parity disk, relieves the write bottleneck that plagues RAID 4. Table 2 summarizes the RAID levels and provides some indication of their respective suitability for particular applications.

Table 1 includes the terms "hot-pluggable," "hot-swappable," and "hot-spareable," which refer to the replacement of a failed drive. The RAID levels that use parity for fault tolerance provide transparent on-the-fly reconstruction of data in the event of a failed drive. In hot-pluggable, hot-swappable systems, you can remove and replace a failed drive without powering down the system. Systems that have a hot-spareable capability incorporate a spare drive that automatically replaces a failed drive after data reconstruction.

Many buses, operating systems

As Table 1 shows, available RAID products support multiple bus interfaces. They also support a variety of buses. As a couple of examples, AdvanSys SCSI host adapters are available for VESA and EISA slots; those from Distributed Processing Technology provide interfaces to ISA, EISA, and PCI buses. Most of the products also provide support for a wide range of operating systems. For example, the AdvanSys host adapters support DOS 4.0 or higher, Windows 3.1 or higher and NT, Novell NetWare 3.12 and 4.01, OS/2 2.1 and higher, SCO Unix 4.2 and higher, and Interactive Unix 3.2 or higher.

Because RAID techniques are appropriate for multiuser network applications, most of the products support such networking systems and software as Banyan Vines and Novell NetWare. Table 1 also shows that the transfer rate for most systems is 10 Mbytes/sec, characteristic of the Fast SCSI-2 host adapters. Some 20-Mbyte/sec products are available, thanks either to duplexing (for example, Megadrive's Power Array) or to the use of Fast and Wide SCSI-2 (for example, Megadrive's MR/10 and Peripheral Land Inc's QuickArray-W). With upcoming PCI-bus products, you can expect faster than 30-Mbyte/sec transfer rates. In the more distant future, look for better than 100-Mbyte/sec Fibre Channel systems.

The terms "software-based" and "hardware-based" in Table 1 also deserve some explanation. Software-based subsystems make use of the host system's CPU for disk-array management chores; the hardware-based systems contain their own dedicated processors. A popular choice is the i960, though some products use an ASIC RISC processor.

You can save some money by purchasing a software-based system, at the expense of speed. Micronet Technology's software- and hardware-based subsystems provide examples. The RISC µP in the hardware system relieves the system CPU of RAID chores and thus doubles the transfer rate, compared with the software system.

Build or buy?

For RAID systems, as with almost every aspect of system design, you face the decision: Roll your own or purchase? Hard-disk drives are relatively inexpensive and becoming more so all the time. If you have a lot of time, you could start from scratch, write your own RAID drive-management code, and configure your own controller and host adapter.

On the face of it, you'd undoubtedly save money, but you'd face a daunting design task. To quote from the release for AT&T/NCR's Series 3 controller, "…the Series 3 controllers are driven by more than a quarter-million lines of embedded RAID software code." Starting not quite from scratch, you could purchase RAID software, such as Paragon 4.0 from BusLogic and configure your own host-adapter and drive-controller hardware. Ref 2 takes you step-by-step through building your own RAID-5 NetWare array, using Paragon 4.0 software.

However, configuring your own RAID subsystem can be scary. Greg Goelv, marketing director for Micropolis, offers the analogy of building a Rolls Royce from a kit rather than buying the car with a tuned engine. Goelv maintains that you're likely to run into trouble if, for example, you want to upgrade your system by adding more disk drives.

Goelv asserts that disk-drive manufacturers often make subtle design changes even for the same drive model, both in firmware and in SCSI code. As a result, you could end up with incompatibility headaches. RAID-subsystem manufacturers, on the other hand, guarantee backward compatibility for their expansion and upgrade products.

Many standards, buses, interfaces, and such go through incompatibility problems; sometimes, vague or ambiguous specs lead to interoperability headaches when you try to mate different manufacturers' products. To avoid such problems with RAID systems, the RAID Advisory Board has done an excellent job of setting standards and specs for RAID technology.

1. The RAIDBook, RAID Advisory Board. $9.95.

2."Building your own RAID-5 NetWare disk array", Chantal Systems, Division of BusLogic Inc.

3."What are disk arrays?" AT&T/NCR.

4."Why disk arrays?" Chantal Systems, Division of BusLogic Inc.