If you are a SQL Server Administrator, eventually you are going to need to request specific storage for your servers. Depending on the setup at your current company, that is all handled in the background by your Storage Administrators, but if you have the power or are asked for your opinion, knowing about RAID (Redundant Array of Independent Disks) technology is important. You can find full technical explanations on the web for this, but I’ll cover the basics from a SQL Server perspective.
RAID uses multiple hard drives to improve availability and/or performance. RAID can overcome I/O bottlenecks that would result from using a single disk, provide resiliency from data loss through mirroring, and remove a single point of failure from a single drive being used.
To understand RAID, there are three terms we need to define first.
Mirroring is an idea you should understand intuitively, but perhaps not the exact details in relation to RAID. Disk Mirroring replicates data from one disk to another, providing a second copy for disaster recovery, and thus requiring at least two disks. RAID mirroring is performed synchronously and mirrored data can be read from either disk. More on Mirroring.
Striping means that the data is being separated onto multiple drives in a consecutive fashion. By separating the data onto different drives, the I/O load can be balanced across the disks and read times are faster. The more disks that the data is striped across, the faster the data throughput will be; however, if one device fails, since the data is spread evenly across all the disks involved in the striping, all the data will be corrupted and unable to be read. More on Striping.
Parity is probably the hardest term to understand or explain. In the most basic, mathematical sense, parity refers to whether an integer is even or odd. For computing, the definition is specifically whether the total value of 1’s occurring in a given binary number is even or odd. For RAID, parity data bits are a combination of existing data bits to provide redundancy. In the case of any single drive failure, the remaining data can be combined with the parity data to reconstruct the missing data. More on parity.
Although there are more levels of RAID, for instance, 2, 3, 4…etc., they are rarely used, especially for SQL Server. I’ll just be explaining the four main types here.
RAID 0 (Striping) This basic form of RAID stripes data across multiple disks. Reads and writes occur simultaneously across all disks involved and thus provides faster reads and writes compared to a single disk. The more disks involved, the faster the performance. This creates multiple points of failure though, and is not really recommended for database use due to the increased vulnerability.
Pros: Improved Performance of Read and Writes
Cons: No Redundancy and any drive failure corrupts all data.
RAID 1 (Mirroring) This RAID level mirrors, or duplicates, data between a minimum of two disks. Mirroring requires 50% more storage since the mirror is an exact copy of the original data. Read speeds are faster since any disk can respond to a read request. Write speeds are reduced due to copying the data to multiple locations. Read times can be as fast as the fastest drive, while write times are often as slow as the slowest drive. If you need a relatively cheap method to protect your data, this is a good option. If one drive fails, you still have a perfect copy of the data on the other.
Pros: Redundancy, Faster Reads
Cons: 50% Extra Storage, Slower Writes
RAID 5 (Striping & Parity) This is the likely the most common type of RAID used, but requires at least three disks. Data and parity calculations are striped across all the disks. Since data is not mirrored, less storage is ‘wasted’, resulting in only a minimum of 1/3rd (1 / total # of drives) of the storage space used for redundancy. In the event of a drive failure, the data can be reconstructed using the parity data, but at a cost. There is a significant impact when one disk fails due to the parity data reconstruction overhead. Losing the 2nd drive in a three disk RAID 5 configuration will result in the entire array going offline and data being lost. Always replace after the first failure as soon as possible! Since write speeds are slower with RAID 5, it is not the best choice for Transaction Logs or Data Files. Backup drives are a prime candidate for this RAID level though since write speed is not as important.
RAID 6 is growing in popularity; it’s identical to RAID 5, except it adds an extra disk with another set of parity data. This RAID level requires a minimum of four disks and can handle up to two drive failures.
Pros: Fault Tolerance, Fast Reads (until a drive fails)
Cons: Slower Write Speeds, More Expensive than RAID 1 or 0
RAID 0+1 (Mirroring & Striping) Data is striped and subsequently mirrored in this RAID level. This incurs the 50% storage redundancy loss. There is fault tolerance for a single drive failure, but this reduces the RAID to essentially RAID 0 with no redundancy.
RAID 10 (Striping & Mirroring) Data is mirrored first and then striped with this method. As long as one drive from each side of the mirror remains, no outage will occur. With more fault tolerance, this is the preferred high-end method over RAID 0+1.
Pros: Fault Tolerance, Best Read/Write Speeds
Cons: Lots of Disks and Lots of Money
Both RAID 0+1 and 10 are hybrid RAID levels and provide the best read and write performance, but double the required storage. This is what everyone wants but not everyone can afford. It combines the best of the previous configurations but requires many disks and therefore a big budget. These RAID levels are best suited for high traffic, like your Data and Log files. If you cannot afford RAID 10, RAID 5 is a decent alternative though.