Managing RAID on Linux

Chapter 1. Introduction

Every system administrator sooner or later realizes that the most elusive foe in sustaining reliable system performance is bandwidth. On one hand, network connectivity provides a crucial connection to the outside world through which your servers deliver data to users. This type of bandwidth, and its associated issues, is well documented and well studied by virtually all system and network administrators. It is at the forefront of modern computing, and the topic most often addressed by both non-technical managers and the mainstream media. A multitude of software and documentation has been written to address network and bandwidth issues. Most administrators, however, don’t realize that similar bandwidth problems exist at the bus level in each system you manage. Unfortunately, this internal data transfer bottleneck is more sparsely documented than its network counterpart. Because of its second stage coverage, many administrators, users, and managers are left with often perplexing performance issues.

Although we tend to think of computers as entirely electronic, they still rely on moving parts. Hard drives, for example, contain plates and mechanical arms that are subject to the constraints of the physical world we inhabit. Introducing moving parts into a digital computer creates an inherent bottleneck. So even though disk transfer speeds have risen steadily in the past two decades, disks are still an inherently slow component in modern computer systems. A high-performance hard disk might be able to achieve a throughput of around 30 MB per second. But that rate is still more than a dozen times slower than the speed of a typical motherboard—and the motherboard isn’t even the fastest part of the computer.

There is a solution to this I/O gap that does not include redefining the laws of physics. Systems can alleviate it by distributing the controllers’ and buses’ loads across multiple, identical parts. The trick is doing it in a way that can let the computer deal seamlessly with the complex arrangement of data as if it were one straightforward disk. In essence, by increasing the number of moving parts, we can decrease the bottleneck. RAID (Redundant Array of Independent Disks) technology attempts to reconcile this gap by implementing this practical, yet simple, method for swift, invisible data access.

Simply put, RAID is a method by which many independent disks attached to a computer can be made, from the perspective of users and applications, to appear as a single disk. This arrangement has several implications.

• Performance can be dramatically improved because the bottleneck of using a single disk for all I/O is spread across more than one disk.

• Larger storage capacities can be achieved, since you are using multiple disks instead of a single disk.

• Specific disks can be used to transparently store data that can then be used to survive a disk failure.

RAID allows systems to perform traditionally slow tasks in parallel, increasing performance. It also hides the complexities of mapping data across multiple hard disks by adding a layer of indirection between users and hardware.

RAID can be achieved in one of two ways. Software RAID uses the computer’s CPU to carry out RAID operations. Hardware RAID uses specialized processors, on disk controllers, to manage the disks. The resulting disk set, colloquially called an array, can provide various improvements in performance and reliability, depending on its implementation.

The term RAID was coined at Berkeley in 1988 by David A. Patterson, Garth A. Gibson, and Randy H. Katz in their paper, “A Case for Redundant Arrays of Inexpensive Disks (RAID).” This and subsequent articles on RAID have come to be called the “Berkeley Papers.” People started to change the “I” in RAID from “inexpensive” to “independent” when they realized, first, that disks were getting so cheap that anyone could afford whatever they needed, and second, that RAID was solving important problems faced by many computing sites, whether or not cost was an issue. Today, the disk storage playing field has leveled. Large disks have become affordable for both small companies and consumers. Giant magnetic spindles have been all but eliminated, making even the largest-drives (in terms of capacity) usable on the desktop. Therefore the evolution of the acronym reflects the definition of RAID today: several independent drives operating in unison. However, the two meanings of the acronym are often used interchangeably.

RAID began as a response to the gap between I/O and processing power. Patterson, Gibson, and Katz saw that while there would continue to be exponential growth in CPU speed and memory capacity, disk performance was achieving only linear increases and would continue to take this growth curve for the foreseeable future. The Berkeley Papers sought to attack the I/O problem by implementing systems that no longer relied on a Single Large Expensive Disk (SLED), but rather, concatenated many smaller disks that could be accessed by operating systems and applications as a single disk.

This approach helps to solve many different problems facing many different organizations. For example, some organizations might need to deal with data such as newsgroup postings, which are of relatively low importance, but require an extremely large amount of storage. These organizations will realize that a single hard drive is grossly inadequate for their storage needs and that manually organizing data is a futile effort. Other companies might work with small amounts of vitally important data, in a situation in which downtime or data loss would be catastrophic to their business. RAID, because of its robust and varying implementations, can scale to meet the needs of both these types of organizations, and many others.