RAID

In 1987, UC Berkeley researchers David Patterson, Garth Gibson, and Randy Katz warned the world of an impending predicament. The speed of computer CPUs and performance of RAM were growing exponentially, but mechanical disk drives were improving only incrementally. As a result, they stated, "We need innovation to avoid an I/O crisis."

The authors famously proposed a solution in their paper, "A Case for Redundant Arrays of Inexpensive Disks (RAID)." They noted that PC disk drives were starting to match the speed of those supplied for mainframes and minicomputers-yet PC drives tended to be better in terms of cost per megabyte. Why? Because of standards such as SCSI, which enabled suppliers to embed functionality that had once required custom controllers.

Looking at what was available on the market, Patterson, Gibson, and Katz concluded that 75 PC disk drives could be lashed together to provide the capacity of a single mainframe drive-with lower power consumption, lower total cost, and 12 times the I/O bandwidth. The snag was Mean Time to Failure (MTTF), clearly much worse for an array of commodity drives than for a bulletproof, silver-plated mainframe unit.

Their paper therefore suggested the use of extra "check" disks, containing redundant information that could be used to recover data in the event of a disk failure. Once a failed disk was replaced , either by a human operator or by electronic switching, data would be reconstructed onto it automatically.

The rest is history, as just about anybody connected with networking is already aware. The I/O problem was widely recognized and, within a couple of years , Intel-based products like the Compaq Systempro (released in 1990) made RAID an expected ingredient in every midrange and high-end server.

So why discuss RAID now? Apart from the fact that we've never published a tutorial about it, the plummeting costs of both disk drives and the circuitry necessary to support disk arrays make RAID more relevant than ever before. It's now affordable for low-end servers and even standalone workstations.

When the first RAID-based products came out, the cost of controllers and capacious SCSI disk drives meant that servers could easily cost $35,000. Vendors were embarrassed by the "Inexpensive" in the RAID acronym and temporarily decided that the "I" stood for "Independent" instead.

Now, with 20Gbyte drives selling for under $200, it's time to be proud of RAID's low cost and leave the misnomer behind. The fact is, there's little that's independent about the drives in a RAID array.

By design, RAID technology hides the characteristics of individual drives from whatever operating system is run on top of it, presenting multiple disks as if they were a single, larger drive. It maps logical disk block addresses to their actual physical counterparts using a variety of algorithms. The differing organizations of data on disk are known as RAID levels, each with its own particular advantages and disadvantages.

Raid 0

RAID level 0 could more correctly be called "AID," because there's no redundancy about it. Data is merely divided into blocks, each one written sequentially to the next drive in the array. If there are four drives in the array, as shown in the figure , each logical I/O is broken into four physical operations.

The point of RAID 0 is performance. Theoretically, it can deliver n times the performance of a single drive, where n is the number of drives in the array. However, tuning the stripe size is important. If it is too large, many I/O operations will fit in a single stripe and take place on a single drive. If it is too small, each logical operation will be broken into too many physical operations, saturating the bus or controller to which the drives are attached.

Reviewers of RAID 0 products on workstations have commented that they offer little advantage with typical applications, such as word processors and spreadsheets. However, in cases where very large files must be opened or saved-on video servers, for example-they can be very beneficial.

Raid 1

RAID 1 is the simplest actual redundant array design, employing mirrored pairs of disk drives. As seen in the figure, it merely creates a duplicate of the contents of one disk drive onto another. While that fact makes RAID 1 easy to implement, it also makes it the most costly (100 percent redundancy) in terms of required disk overhead.

RAID 1's write performance is slower than that of a solo drive, since all data must be written twice. However, buffering on a controller usually hides this fact from the host computer.

Striping, Mirroring, and Parity. RAID uses a variety of methods to organize data on a disk.

Reads can be faster, since it's always possible to retrieve data from whichever drive is available sooner.

Raid 2

RAID 2 is a bit-oriented scheme for striping data. Each bit of a data word is written to a separate disk drive, in sequence. Checksum information is then computed for each word and written to physically separate error-correction drives.

Unfortunately, I/O is slow, especially for small files, because each drive must be accessed for every operation. Controller design is relatively simple, high data-transfer rates are possible for large files, and disk overhead is typically 40 percent. However, while reliable, RAID 2 is seldom considered worth bothering with today.

Raid 3

RAID 3 introduces a more efficient way of storing data while still providing error correction. It still stripes data across drives bit by bit (or byte by byte). However, error-checking now takes place by storing parity information (computed via a mathematical function known as the Exclusive OR, or XOR) on a separate parity drive (see figure).

Given that parity values are simple to compute and write, RAID 3 arrays can perform swiftly. However, any I/O operation must address all drives simultaneously . This means that, while RAID 3 delivers high data-transfer rates, it is best suited to large files such as video streams.

Raid 4

RAID 4 modifies the RAID 3 concept by working with data in terms of blocks (as does RAID 0), rather than bits or bytes. This reduces processing overhead and can make for high aggregate data-transfer rates on reads. For writes , however, there is inevitable contention for the sole parity drive, making this RAID level relatively sluggish .

Raid 5

One of the most popular RAID levels, RAID 5 is again block-oriented and based on the storing of parity information. However, instead of placing parity data on a single drive, it distributes it across the entire array (see figure).

Because RAID 5 eliminates the parity-drive bottleneck, it enhances write performance. And due to the independence of all the drives in the array, read performance is tops among true RAID levels. Recovery following a disk failure is relatively slow, but reliable enough. All in all, RAID 5 achieves an excellent balance between performance, data protection, and low cost.

RAID 10 And RAID 53

RAID 10 is also known as RAID 0+1 or 1+0 because it combines the elements of RAID 0 and RAID 1. It uses two sets of drives that mirror one another, as in RAID 1. Then, within these sets, data is striped across the drives (as in RAID 0) in order to speed access.

RAID 53, which should really be called RAID 30 using the above logic, combines RAID 0 and RAID 3. Again, it uses a striped array, as with RAID 0, but the segments of this are RAID 3 arrays. High data-transfer rates and high I/O rates for small requests are both offered -but at a price.

Enhancing Performance

RAID controllers can become a bottleneck, especially with high-speed interconnects like Fibre Channel, because of the calculations they must perform. For example, to perform a disk-write operation to a RAID 5 array, a Read-Modify-Writeback operation must be performed. First, old data must be read from both a data drive and a parity drive. Second, that data must be XORed. Third, new data must be written to the data drive. Fourth, new data must also be XORed with the parity data, and only then can the result finally be written to the parity drive.

One solution to this bottleneck has been to move the responsibility of calculating XOR data to the disk drives themselves . Seagate, IBM, and other vendors have released drives that can perform XOR calculations in parallel with other disks, without the aid of the RAID controller.

The industry is entering a period of rapid transition in I/O architectures. InfiniBand's 2001 products will couple I/O directly to host memory, offering transfer rates of up to 6Gbytes/sec, and RAID products will evolve to support such throughput. At the same time, the falling cost of controllers and drives will make RAID arrays ever more commonplace on the low end. Your next notebook computer may even offer you a choice of RAID levels.

Resources

"A Case for Redundant Arrays of Inexpensive Disks (RAID)" by Patterson, Gibson, and Katz is archived at http://sunsite.berkeley.edu/Dienst/UI/2.0/Describe/ncstrl.ucb/CSD-87-391/.

Details on Storage Computer's proprietary RAID 7 can be found at www.raid7.com/wp__raid7afa.html.

The RAID Advisory Board Web site is located at www.raid-advisory.com.

A search engine designed specifically to find information about RAID and other storage- related topics can be found at www.searchstorage.com.

This tutorial, number 144, by Jonathan Angel, was originally published in the July 2000 issue of Network Magazine.