Practical RAID Performance?

Question

I've always thought the following to be a general rule of thumb for RAID:

• RAID 0: Best performance for READ and WRITE from stripping, greatest risk
• RAID 1: Redundant, decent for READ (I believe it can read from different parts of a file from different hard drives), not the best for WRITE
• RAID 0+1 (01): combines redundancy of RAID 1 with performance of RAID 0
• RAID 1+0 (10): slightly better version of RAID 0+1
• RAID 5: good READ performance, bad WRITE performance, redundant

IS THIS ASSUMPTION CORRECT? (and how do they compare to a JBOD setup for R/W IO performance)

Are certain practical RAID setups better for different applications: gaming, video editing, database (Acccess or SQL)?

I was thinking about hard disk drives but does this apply to solid state drives as well?

Solution 1

Raid 0. Great read and write speed. Failure risk increased as the number of member disk increase. No parity.

Raid 1. Great read speed only if driver properly implemented - If you use Areca and LSI raid controllers, they can deliver almost the same read capability for Raid 1 sets as Raid 0 sets (within 10%). Note that for software raid solutions there are two types, OS-software and Motherboard-software.

Most motherboard-type Raid does not offer good Raid 1 read performance. Last time I checked Windows and Linux did not have a good implementation of Raid 1 reads too. BSD had a correctly implemented Raid 1 implementation that uses roundrobin read method.

In short, use Raid 1 for redundancy, and for read speed if you uses advanced controllers.

Raid 1/0 & Raid 0/1. This is a combination thing. Let's say you have 10 disks. Raid 1/0 is that there is mirror in each set: Set A (1+2), Set B (3+4), Set C (5+6), Set D (7+8), Set E (9+10), then Do a strip set across the 5 sets (A-E). This way, each of the set can have one drive failure, but if two drives fail in the same set you are done.

Raid 0/1 is that to have strip set A (1+2+3+4+5) and set B (6+7+8+9+10) and mirror the two sets. This way, if drive 2 and drive 9 fails, most controllers do consider this as a total failure (which, in fact, you still have all the data).

There is little performance difference between the two but I don't think you usually get both from a Raid controller.

Raid 5. a very mixed bag: for sequential reads, it is faster than raid 1/0, and for random read it is slightly slower that that. Note that the performance of raid 5 is very dependent on the speed of the controller (e.g. you can't expect much in onboard raid).

Raid 6. Redundancy increased compared with Raid 5. two drive can fail at anytime, and when rebuilding the array after 1 drive failed there is still redundancy (note that when raid 5 drive failed, the array is similar to raid 0 - any drive fail = total loss).

JBOD. No advantage that i can think of.

Solution 2

As your question implies, there is not really a "best" RAID configuration, only best for a particular set of circumstances, with cost often being one of the most important factors.

Without going into the minutiae of controllers and software, here would be my rules of thumb.

RAID 0 is the fastest since you can read from and write to many disks at once, and no space is "wasted" for redundancy. Lose any disk and you lose the set, so RAID 0 should only be used on a machine you do not care about, or is easy to restore and does not contain data you value. A gaming machine might fit this scenario, though honestly, the speed difference is not so noticeable that I would be willing to accept the increased risk of having to rebuild the machine. It can also be useful for really fast "scratch" areas if you have software that needs that sort of thing.

RAID 1 is the very common mirror setup, usually of 2 disks. Reads are generally just shy of twice as fast as a single disk, while writes are just a bit slower than a single disk. On servers, RAID 1 is an excellent choice for your operating system files. It is also a good choice when you need redundancy, your storage needs are not so great as to require RAID 5 and you might benefit from the extra read speed (database log files are commonly placed on RAID 1).

RAID 2 through 4 (not in the question) are generally unused except by certain vendors' products on the the enterprise side.

RAID 5 is a compromise between not wasting too much space on redundancy and still getting the added performance of extra disks. Read speeds are very good since all the disks can participate. Write speeds can sometimes be a problem with RAID 5, though I think this is sometimes overstated depending on the situation. Even with hardware that does the parity calculations, small random writes suffer on RAID 5 since each logic write operation requires 4 I/Os (read from data disk, read from parity disk, write data disk, write parity disk). Pick RAID 5 when you want to maximize the amount of storage you get out of a disk set while still having a good degree of safety. Avoid when your application has high performance needs and requires lots of small random writes (virtual machine hard disks, database data files). Also be aware that modern large disks take a loooooong time to rebuild when they fail as part of a RAID 5 set, and that puts your data at risk longer for a second disk failure. RAID 6 can reduce this risk at the expense of even worse random write performance.

RAID 1/0 and 0/1 have pretty much the same performance characteristics. 1/0 is preferable though because a failure means rebuilding only the pair of disks involved rather than an entire stripe of disks. RAID 1/0 is the fastest general purpose configuration of the ones mentioned in the question. Read and write performance are both great (reads can happen from all disks at once essentially, writes have to happen to multiple disks but no read/read/write/write cycle needed like RAID 5) though RAID 5 can win in certain circumstances. RAID 1/0 is (unsurprisingly) also the most expensive in nearly all cases. It should only be used when performance is critical, the data is highly valuable, and the application has no tolerance for down time. Often database servers (or at least their data files) make good candidates for RAID 1/0. I also prefer to put virtual machines on RAID 1/0 if there are going to be a lot of them or they run disk intensive applications that I really care about.

Solid State Drives can indeed be put into RAID configurations, particularly for its redundancy value. Many SSDs can outperform the bus (SATA) they are attached to though, especially for read operations, so the performance effect of RAID is less compelling with SSDs.

Comments

[wag2639]

My understanding of the redundancy of each of the RAID classes is pretty good, I'm just unsure of the read and write I/O performance.

[wag2639]

As for JBOD, I know it doesn't offer performance or redundancy, I just mentioned it as a base reference.

[user1984103]

JBOD can't stripe usually because it's designed for use with disks of different sizes - if you have the same size, you should already be using RAID 1 or RAID 0 instead.

[wag2639]

thanks.. so from what i'm reading, using general consumer motherboards with the goal of performance gaming, raid 0 is best without getting a separate dedicated raid controller? (i don't care if i have to reinstall os + programs and data will be elsewhere, possible raid 5 array)

[Hans Kilian]

you are very correct. though i am afraid that the performance using motherboard raid won't be very different from using OS-based software raid (though, of course, the difference is that your system drive cannot be raided in OS-based software raid). Also, raid5 performance in motherboard raid is pathetic (downward of 15-20MB/s in some cases!)