Since I’ve been doing a lot of coverage of storage technology both for the enterprise and for the
home lately, I thought I should give an explanation of what RAID storage is. I won’t go in to
every RAID type under the sun, I just want to cover the basic types of RAID and what the
benefits and tradeoffs are.
RAID was originally defined as Redundant Array of Inexpensive Drives, but RAID setups were
traditionally very expensive so the definition of “I” became Independent. The costs have recently
come down significantly because of commoditization and RAID features are now embedded on
to most higher-end motherboards. Storage RAIDs were primarily designed to improve fault
tolerance, offer better performance, and easier storage management because it presents multiple
hard drives as a single storage volume which simplifies storage management. Before we start
talking about the different RAID types, I’m going to define some basic concepts first.
Fault tolerance defined:
Basic fault tolerance in the world of storage means your data is intact even if one or more hard
drives fails. Some of the more expensive RAID types permit multiple hard drive failures without
loss of data. There are also more advanced forms of fault tolerance in the enterprise storage
world called path redundancy (AKA multi-path) which allows different storage controllers and
the connectors that connect hard drives to fail without loss in service. Path redundancy isn’t
considered a RAID technology but it is a form of storage fault tolerance.
Storage performance defined:
There are two basic metrics of performance in the world of storage. They are I/O performance
and throughput. In general, read performance is more valued than write performance because
storage devices spend the majority of their time reading data. I/O (Input/Output) performance is
the measure of how many small random read/write requests can be processed in a single second
and it is very important in the server world, especially database type applications. IOPS (I/O per
second) is the common unit of measurement for I/O performance.
Throughput is the measurement of how much data can be read or written in a single second and it
is important in certain server applications and very desirable for home use. Throughput is
typically measured in MB/sec (megabytes transferred per second) though mbps (megabits per
second) is sometimes also used to describe storage communication speeds. There is sometimes
confusion between megabits versus megabytes since they sound alike. For example, 100 megabit
FastEthernet might sound faster than a typical hard drive that gets 70 MB/sec but this would be
like thinking that 100 ounces weighs more than 70 pounds. In reality, the hard drive is much
faster because 70 MB/sec is equivalent to 560 mbps.
RAID techniques defined:
There are three fundamental RAID techniques and the various RAID types can use one or more
of these techniques. The three fundamental techniques are:
Mirroring
Striping
Striping with parity
Mirroring:
Data mirroring stores the same data across two hard drives which provides redundancy and read
speed. It’s redundant because if a single drive fails, the other drive still has the data. It’s great on
read I/O performance and read throughput because it can independently process two read
requests at the same time. In a well implemented RAID controller that uses mirroring, the read
IOPS and read throughput (for two tasks) can be twice that of a single drive. Write IOPS and
write throughput aren’t any faster than a single hard drive because they can’t be process
independently since data must be written to both hard drives at the same time. The downside to
mirroring is that your capacity is only half of the total capacity of all your hard drives so it’s
expensive.
Striping:
Data striping distributes data across multiple hard drives. Striping scales very well on read and
write throughput for single tasks but it has less read throughput than data mirroring when
processing multiple tasks. A good RAID controller can produce single-task read/write
throughput equal to the total throughput of each individual drive. Striping also produces better
read and write IOPS though it’s not as effective on read IOPS as data mirroring. You also get a
large consolidated drive volume equal to the total capacity of all the drives in the RAID array.
Striping is rarely used by itself because it provides zero fault tolerance and a single drive failure
causes not only the data on that drive to fail, but the entire RAID array. Striping is often used in
conjunction with data mirroring or with parity.
Striping with parity:
Because striping alone is so unreliable in terms of fault tolerance, striping with parity solves the
reliability problem at the expense of some capacity and a big hit on write IOPS and write
throughput compared to just data striping. Data is striped across multiple hard drives just like
normal data striping but a parity is generated and stored on one or more hard drives. Parity data
allows a RAID volume to be reconstructed if one (sometimes two) hard drives fail within the
array. Generating parity can be done in the RAID controller hardware or done via software
(driver level, OS level, or add-on volume manager) using the general purpose processor. The
hardware method of generating parity either results in an expensive RAID controller and/or poor
throughput performance. The software method is computationally expensive though that’s no
longer a problem with fast multi-core processors. Despite the performance and capacity penalty
of using parity, parity uses up far less capacity than data mirroring while providing drive fault
tolerance making this a very cost-effective form of reliable large-capacity storage.
