How do RAID arrays scale as you increase the number of hard drives they contain? Part 1 of our RAID Charts project shows all the benchmark results for RAID 0, RAID 1 and RAID 0 1 setups across two to eight disk drives.

I notice that at a queue level of one, The results of all configurations are approximately the same. As a single user in a desktop environment, I expect my queue depth to be close to one or two, never 64, giving me very little benefit from multiple devices. I would have liked to see in the study a base line of a single non-raided device of the type used in the study, as well as a single raptor150. I would also like to have seen a benchmark workload that would represent what a single user would be doing.

we have a baseline... it was the RAID 1 benchmark, RAID1 only performs as fast as a single drive...

PS, i noticed that an even number of drives in the RAID 0 configuration provided better performance than that of odd drives, anyone know any particular reason towards this? And also, just wondering, what was the stripe size, it can generally vary between 4K and 256K (higher stripe size for bigger sequential read/writes, lower stripe size for faster access times, 64K is usually a happy middle.)?

No... RAID 1 performs slower than a single drive, at least as far as writes are concerned (not so sure about the effects on other tests). A baseline would have been a single drive, the test provided no baseline within the article that I could find.

I agree I would have liked to have seen the single drive results at a minimum to give some relativity to the results.

With the controller used I would say that writes would be slightly slower and the reads slightly higher, RAID1 speed is highly dependent on the controller used.

I too was looking for a single drive baseline

If (big if) the controller is doing it's proper job at its proper speed, it won't influence the speed and will run just as fast as a single drive, because the controller sends and receives the data from one drive and its cache as well as sending the same data to another "mirrored" drive, but instead of reading the data back, it just does some verifying, now, if the controller is working properly this should not effect performance, however, this is generally not the case, and the verifying takes a bit of the controllers time and slows the write speeds down a bit because it has to happen to both drives, one that is idle and one that is verifying, it doesn't effect reads though because that only happens to the idle drive, so reads seem faster but aren't really.

Raid 5 is the most popular raid level so why were there no raid 5 tests?

JT001 is correct that RAID 1 speeds are highly controller dependent.

I've seen RAID 1 controllers that are slightly slower than a single drive on reads, and an appreciable amount slower than that on writes.

I've also seen high-end RAID 1 controllers from LSI that are equal to a single drive on writes, and faster than a single drive on reads. (This is possible because the controller intelligently interleaves reads to both drives, taking advantage of the fact that 2 copies of the data exist).

So you can't make blanket statements that RAID 1 is always slower because of this or that. The design of different controllers and their manner of operation translate into a wide variety of different RAID 1 performance levels.

Ummm...

I agree That there should be a single drive benchmark, especially for the IO test profiles. This can't even be found in the HDD Charts, as only the HD501LJ and HD300LJ is listed. The HD321KJ is not, and the different model number would tell me that it's vital statistics would be different???

I hope the final part of this article also delves into the statistical probabilities of data loss with each RAID array. That's a necessary part of the calculus of a cost - risk analysis. What's been presented so far is certainly interesting but I want to be able to decide how much statistical risk (of data loss) I'm assuming for each configuration...and then I can make a rational decision on how much I'm willing to spend.

Of course, you'll have to start with some assumptions about statistical hard drive failure rates and drive enclosure rates but as long as those are realistic and standardized across all calculations then you'll be able to give us relative indicators of failure probability...obviously not absolute predictions.

Disk 4's going to drop out. Next Wednesday, around Teatime.

This 3 year old article will give you the RAID-5 part:
http://translate.google.com/transl [...] de&ie=UTF8

It's kind of ridiculous that with the "current" hardware the access time goes up to 30ms, while it did not with 8 drives in more computation intensive RAID-5 three years ago I would have thrown the controller away and gotten a faster one.

Since Patrick Schmid was also part of the testing crew that time, I'm wondering why he published this ...

I don't think the test bed was properly configured for this observation you pointed out (which is NOT true for 3 drives, an odd number). 3 drives is always faster than 2 and less than 4.

Now, it is interesting that 7 drives is always slower than 6 and 8.

Then again, past 4 drives, you've reached the saturation of the controller. I guess the thing to learn is don't oversaturate your controller.

You missed one very important difference between RAID 0+1 and 1+0. Given the two, RAID 1+0 is always preferable because the rebuild time is significantly lower than with RAID 0+1.

In RAID 1+0 (stripe of mirrors), a failure of a disk results in only the mirrored-pair being degraded. When the disk is replaced, only the data on the mirrored pair must be rebuilt.

In a RAID 0+1 (mirror of stripes), a failure of a disk places the entire stripe in a degraded state. When the disk is replaced, the data against the entire stripe must be re-mirrored against the other functional stripe.

So, given the choice, it is always better to run RAID 1+0 over RAID 0+1.

I am putting together a lab for developing in Linux\Unix and there aren't very many performance reviews for Raid controllers in Unix. I have heard of a lot of problems with adeptec and highpoint support on these platforms. 3ware is the most solid I have heard.

A good article review of Raid controllers in Unix/Linux with performance numbers and driver maturity would be very helpful.

As tostada said, you _always_ want to do 1+0. Fortunately, many controllers that _say_ they're doing 0+1 _really_ do 1+0. The only way to tell for sure is to intentionally fail a drive and then watch the blinky lights on a rebuild.

As someone else mentioned, RAID 1/1+0/0+1 on n disks allows for up to n times the read performance of one disk, and up to n/2 times the write performance. Sequential access will show the most benefit, as the latency increase won't matter.