How to choose a RAID level for databases, VMs, file servers and backup targets

The right RAID level depends on what the server does. A database that writes small random blocks all day needs something different from a backup repository that writes large sequential files at night. This guide turns the RAID guidance in the Dell PERC and HPE Smart Array / MR controller guides into a practical decision for the workloads we see most often on PowerEdge and ProLiant servers.

What the vendors recommend

HPE's controller guide says the RAID choice rests on three needs: the fault tolerance required, the write performance required and the usable capacity you need. It then maps levels to those needs:

• RAID 1/10 Triple: optimise for fault tolerance and write performance.
• RAID 6/60: optimise for fault tolerance and usable capacity.
• RAID 1/10: optimise for write performance.
• RAID 5/50: optimise for usable capacity.

Dell's PERC 11 guide gives workload examples: RAID 1 for a simple boot disk, RAID 5 or 6 for file or web servers (sequential reads and writes of files) and RAID 10 for transactional databases (small random reads and writes).

Decision table by workload

Write cost decides most database and VM designs

HPE's guide lists the disk I/Os caused by every host write: 1 for RAID 0, 2 for RAID 1/10, 3 for RAID 1/10 Triple, 4 for RAID 5 and 6 for RAID 6. Suppose a database needs 2,000 random writes per second. On RAID 10 the drives must absorb about 4,000 I/Os per second, on RAID 5 about 8,000 and on RAID 6 about 12,000. To reach the same write rate with parity RAID you need more or faster drives.

A controller with protected write cache can absorb bursts and combine small writes, which reduces the penalty in practice. How that works, and why the cache needs a healthy battery or energy pack, is explained in write-back vs write-through cache.

Databases and transactional applications

• Use RAID 10 as the default. HPE describes it as the highest write performance of any fault-tolerant configuration, and it survives several failures as long as no two failed drives are in the same mirrored pair.
• Where the controller supports it (HPE Smart Array P-class SR Gen10, SR Gen10 Plus and SR Gen11), RAID 10 Triple keeps three copies. HPE calls it the highest data protection of any configuration, at one-third usable capacity.
• RAID 10 needs an even number of drives; plan capacity as half of raw.

Virtualisation hosts

Dell's PERC 11 guide warns that hypervisor configurations running several workloads on one RAID array may see degraded performance or latency, because each VM's I/O reaches the array as random I/O. Dell suggests running fewer workloads per array, or separate arrays and physical disks for each workload, and keeping write-back with read-ahead on rotational disks or using SSDs for random I/O. Background operations such as initialization, consistency checks and reconstruction also reduce performance while they run.

In practice this means separating the hypervisor boot from the datastores (see boot-optimized drives), and placing database VMs on RAID 10 while general-purpose VMs can sit on a larger RAID 6 set. For sizing the rest of the host, see server sizing for virtualization.

File servers, web servers and backup targets

These workloads favour capacity, so parity RAID fits. Two vendor statements tip the choice toward RAID 6 on large drives:

• Dell: RAID 6 protects against two drive failures and against a failure while a single disk is rebuilding.
• HPE: a RAID 5 or RAID 6 rebuild generally takes about 15 to 30 seconds per gigabyte, depending on I/O load, drive count, rebuild priority and drive performance. Large drives therefore stay degraded for a long time. See rebuild time and risks.

HPE also states that the fault tolerance of RAID 5, 50, 6 and 60 decreases as the array gets larger. For very large repositories, RAID 60 with more parity groups raises fault tolerance at the cost of capacity.

A backup repository on RAID 6 is still only one copy. Keep at least one more copy elsewhere, as in the 3-2-1 strategy.

Check the controller and drive count

• Entry controllers support fewer levels. For example, the Dell PERC H355 and H350 support only RAID 0, 1 and 10. The full list per controller family is in RAID levels explained.
• RAID 5 needs 3 drives, RAID 6 needs 4, RAID 50 needs 6 and RAID 60 needs 8. HPE requires the drive count in RAID 50/60 to divide exactly by the number of parity groups.
• Keep one drive type per array: Dell and HPE do not support hard drives and SSDs in the same virtual disk or array.

Frequently asked questions

Is RAID a backup?

No. RAID keeps a volume online through drive failures, but it copies deletions and corruption instantly. HPE's guide tells you to make sure the array has a current, valid backup before replacing drives.

RAID 5 or RAID 6 for a file server with large hard drives?

RAID 6, if the controller supports it. Dell states that with a single array RAID 6 is more effective than a hot spare, and HPE's rebuild estimate shows how long a large drive stays exposed during a RAID 5 rebuild.

Checklist

• Classify each workload as random or sequential, and as performance-first or capacity-first.
• Databases: RAID 10. File, web and backup: RAID 6 (or RAID 5 on small arrays). Scratch: RAID 0.
• Calculate the disk I/O needed with the write cost (2, 4 or 6 per write).
• Separate boot from data, and busy VMs from general VMs where possible.
• Confirm controller support and drive-count rules before ordering drives.
• Keep backups outside the array.

Skyline can supply, configure and maintain Dell, HPE and Huawei servers under an annual maintenance contract (AMC). Contact us.