400GB Hard Disk Drives in RAID 0, RAID 5 and RAID 10 Arrays: Performance Analysis

Today we are going to conduct a detailed study of RAIDability of contemporary 400GB hard drives on a new level. We will take two "professional" drives from Seagate and Western Digital and four ordinary "desktop" drives for our investigation. The detailed performance analysis and some useful hints on building RAID arrays are in our new detailed article.

by Nikita Nikolaichev
02/23/2007 | 08:17 AM

Four years ago I tried to compare the RAIDability of a few versions of firmware for then-popular Seagate Barracuda ATA IV hard disk drives (for details see our articles called Seagate Barracuda ATA IV in RAID 0: Truth and Fiction and Seagate Barracuda ATA IV HDD Review). Those tests produced discouraging results because the drives, performing in a RAID0 array, would be slower in some tests than a single such HDD. That article proved to be a bit controversial and I decided to carry out such tests in the future for my personal use only, without publishing them.


But there is now a definite trend on the HDD market to produce “professional” versions of desktop HDDs for use in entry-level servers, data storage systems and workstations. Such HDDs are endowed with special technologies that should ensure required performance at a much higher reliability level than with ordinary HDDs. The drive’s operation under vibration is particularly paid much attention to.

So, in this article I’ll return to the topic of RAIDability of today’s HDDs on a new level. Testing HDDs of a particular model from a particular company in a RAID1 array wouldn’t tell me what influences the outcome more, the HDDs or the RAID controller employed. It’s only through comparing different HDDs on the same controller that any practical conclusion can be drawn from the test. Of course, these practical results are going to be bound to the controller the test was performed on.

As Archimedes once said, you cannot move the Earth without having a place to stand. Choosing a specific controller for such a place, I lose one degree of freedom, but instead acquire the opportunity to compare hard disk drives. And that’s exactly what I want.

So, what about the controller? Not long ago the Areca ARC1220 controller was tested on our site and it’s going to suit me fine judging by its excellent results in our tests. Next, the types of arrays and the number of drives in them are to be decided upon. After some deliberation I decided to check out RAID0, RAID5 and RAID10 arrays made out of four HDDs.

Here are the HDD models to be tested on the Areca ARC1220 controller:

This list includes two “professional” drives from Seagate and Western Digital (ST3400832NS and WD4000YR) and four ordinary “desktop” drives. To build the arrays I used four samples of the same HDD model, from the same batch and with the same firmware version.

I chose 400GB drives due to their best capacity/price ratio. If you don’t see your favorite HDD model among these, I can assure you that you’ll see it in one of our upcoming reviews. We are going to make such reviews on a regular basis.

But you may be wondering if this test is at all necessary? We have recently tested modern large-capacity HDDs on our site and know how things stand among them (for details see our article called 500GB HDD Shootout: Seagate Barracuda 7200.10 and Others!). Is that not enough? It is not because I’ve got somewhat different models today and, furthermore, HDDs behave differently in a RAID array than on their own.

On one hand, the average load on each drive is lower in a RAID array because the request queue is shared by all the drives within the array. But on the other hand, each request to the array may transform into several requests to the drives, depending on the array type and the sophistication of the RAID controller’s firmware. For example, a request to write to a RAID5 array is, in theory, transformed on the controller into two read requests and two write requests, sent to two different HDDs. Or taking a RAID1 array as an example, the controller doubles each read request it receives and sends the two resulting requests to both HDDs in a maniacal desire to make sure that the contents of both are identical.

This affects the timing characteristics of load, I mean the time between two sequential requests to the same HDD. And it requires the drive to change its anticipatory reading and deferred writing algorithms accordingly.

And finally, the disk load is also influenced by the controller’s caching algorithms (especially if the controller is equipped with its own cache memory).

That’s enough, though. Let’s get to business now.

Intel IOMeter Database Pattern

You’ll now see the drives perform under a load typical of database servers. This IOMeter pattern is sending requests to read and write 8KB data blocks. Throughout the test, the outstanding request queue is getting deeper and the percentage of reads to writes is changing, too.

For each array type there are three diagrams that show the dependence of performance on the percentage of write requests at three different loads (1, 16 and 256 outstanding requests).

The Hitachi drive gains the lead in a RAID0 array under low load, but it is outperformed by the Western Digital HDDs as the percentage of writes grows up. In the Random Write mode (i.e. at 100% writes), the Maxtor suddenly leaps to first place, although it shares last places with the Seagate drive at lower percentages of writes.

At higher load the WD drives are ahead of the Hitachi from the very start. Especially good is the WD4000KS model which is in the lead most of the time. It only proves to be slower than the previous generation of WD drives at some percentages of writes.

It is the Seagate HDD that occupies last place, just like in the previous case.

The WD4000KS is just immodestly superior to the other HDDs at this load. I guess its success is indicative of its NCQ support, but the WD website keeps silent on that subject.

Note also the small success of the Seagate drive, but it is indeed small in comparison with the WD4000KS.

Let’s now see if anything’s different when the HDDs work in RAID5.

Here is a good illustration of how the controller’s algorithms may affect performance of a RAID array. Note the jagged shape of the graphs – this is the consequence of a conflict between the controller’s algorithms and the HDDs’ deferred writing.

Note also that the Hitachi is not as good as in RAID0, but the drives from WD are on top again, the WD4000KS still being the best of them.

The graphs take a familiar look at higher load – a gently sloping line that is going down steadily as the percentage of writes is growing up.

The WD drives are in the lead again, and the Hitachi is only competitive against them when there are few writes to be done. The two outsiders in the bottom part of the diagram are the drives from Maxtor and Seagate.

The WD4000KS leaves its opponents far behind at the highest load while the Seagate team outperforms the Maxtors.

Now, the last thing left to do is to check the HDDs out in RAID10.

We seem to have got some fighting here! The Hitachi looks good at low load again and is competitive against the WD4000KD and WD4000YR. The WD4000KS doesn’t perform too well here, but it is the NL35.2 from Seagate that takes last place.

The WD4000KS perks up at higher load. It is the leader when there are few writes to be performed. The other two HDDs from Western Digital are in the leading group, too.

The WD4000KS is again a little faster than the others at high percentages of reads (due to NCQ, I suppose).

So, this round is won by the WD drives. The ability of the WD4000KS to speed up at high loads should be emphasized specifically.

Intel IOMeter File-Server & Web-Server Patterns

These patterns emulate the typical load on the disk subsystem of a file- or web-server. The File-server pattern comes first.

The drives split up into three groups while working in RAID0 arrays. The first group includes the WD4000KS all alone. This HDD enjoys a colossal advantage over its opponents at high loads.

The Hitachi and the two other models from WD are contending in the second group. It’s hard to tell the winner among them as the Hitachi is faster at low loads and the WD drives at high loads.

The third group includes the HDDs from Seagate and Maxtor. No comments about these two.

The WD drives seem to work better than the Hitachi in RAID5. The three HDDs from WD are the leaders here.

The WD4000KS doesn’t feel at ease in RAID10 at low loads. But as the load grows higher, this model overtakes the other HDDs from Western Digital and claims leadership for itself.

There are no write requests in the Web-server pattern, so I guess the WD4000KS is going to be in its element there.

That’s right. The Hitachi is fast at low loads, but the WD4000KS proves its superiority eventually.

We’ve got the same picture with the other array types. The WD4000KS wins this round.

Intel IOMeter Workstation Pattern

These patterns emulate applications working with the NTFS file system. There’s a large share of write requests here, so HDDs other than the WD4000KS have got a chance :).

Well… The drives from Western Digital are still in the lead. The two older models are the leaders at low loads, and the new model at high loads.

In RAID5 mode the WD4000KS overtakes the older drives from WD at even lower loads than in RAID0. Is it the result of a “virtual load increase”?

The picture for RAID10 mode doesn’t differ much from what you’ve seen in RAID5 mode.

It is also interesting to know how the results may be affected by the reduction of the test’s operating zone to 32 gigabytes.

As you can see, the Hitachi drive is very sensitive to the reduction of the test zone. This drive has fewer heads than the others, and this reduction narrows the platter zone the heads are moving within. Thus, the chance of a short seek is increased and the average time to process a read or write request is reduced.

Well, this helps the Hitachi in its struggle with the HDDs from WD at low loads only.

Hitachi’s secret weapon doesn’t work with RAID5. I guess that it is more important to have aggressive deferred writing algorithms here.

It’s strange to see the Maxtor lagging behind. It seems to be the only one not to benefit from the reduction of the test zone. Or perhaps it just doesn’t like to work in RAID5?

In RAID10 mode the Hitachi drive leaves the WD4000KS behind at low loads, but that’s all it can do. The Maxtor looks much better in RAID10 than in the previous case.

Summing it up, the HDDs from Western Digital have been superior in the workstation tests, too.

Intel IOMeter Sequential Read & Write Patterns

These patterns help investigate the performance of the controller and HDDs at sequential reading and writing. IOMeter is sending a stream of read/write requests with a queue depth of 4 to the array. Each minute the size of the requested data block is changed so that we could see the dependence of the sequential read/write speed on the data block size.

So, we’ve got an opportunity to give some praise to the Seagate drive. Without a doubt, it is superior in this pattern. The Hitachi team has the lowest results because it is the least dense of all the HDD models included into this test.

The maximum read speed is reached sooner by the RAID5 arrays than by the RAID0 ones – on 4KB and 8KB data blocks, respectively. Note the way the WD4000KS-based array is reaching its max speed – its graph is smoother than the others.

But the WD4000KS model does much better in a RAID10 array and outperforms all other HDDs. Let’s see now what we have at writing.

The WD4000KS is best in this test. Its large cache and high areal density put it ahead of its opponents.

We’ve got another fit of “sea sickness” in RAID5 mode. None of the arrays could draw a flat graph. The controller’s operation algorithms in RAID5 mode seem to be exceedingly sensitive to the size of the data block and the HDDs’ segmentation at deferred writing.

Curiously, the Maxtor-based array did well in this test!

The WD4000KS is again among the leaders in RAID10 mode.

Intel IOMeter Multithreaded Read & Write Patterns

This is an exciting test, too. Its point is in reading from (or writing to) the array in several threads. Our tests on single HDDs have revealed a few interesting facts. First, Maxtor’s drives have been good irrespective of the number of threads. Second, Seagate’s drives have been very poor in the multi-threaded reading test. But this time I’m testing somewhat different drives from Seagate, and in an array, and on an Areca ARC1220 controller. It means more variables are added into the equation – let’s solve it!

So, the RAID arrays perform roughly in the same way as the corresponding single HDDs.

Maxtor is still the king of this test. I am also pleased with the results of the Hitachi drive when processing several threads because it is no worse than the drives with much higher areal density. The Seagate drive does well, too. At least you don’t see such a horror as in the review of the 7200.10 series. And once again in this article I can’t but praise the WD4000KS drive. It shows how far WD’s programmers have progressed with the transition to the new generation of HDDs.

The speeds go down a little in RAID5 mode, but the overall picture remains unchanged.

It’s the same with RAID10: the Maxtor is on top, followed by the WD4000KS.

Let’s now see what we have at writing.

There’s nothing much exciting at writing. All the drives make use of deferred writing and all of them do it right. The Hitachi has a smaller cache buffer than its opponents and this affects its results, which are lower, even though not exactly by half.

The WD4000KS gains the lead in this mode.

And the same WD4000KS is also the best in its favorite RAID10 mode.

So, we’ve got two heroes in this test: the Maxtor shows its excellent abilities of multi-threaded reading in RAID arrays, too. And the WD4000KS is the best of all at writing.

FC-Test 1.0

Our favorite File Copy Test is the last on our test program.

We create two logical volumes on the arrays, 32GB each, and format them in NTFS and then in FAT32. Next, we create a file-set on the first volume, then read it, copy it into a folder in the same volume (Copy Near), and finally into a folder in the second volume (Copy Far).

The first test brings a surprise already: when united into RAID0 or RAID10 arrays, the HDDs from Maxtor and Seagate are not very successful at creating small files. The Hitachi drives are surprisingly weak as a RAID5 array. The only drive I can find no fault with is the WD4000KS!

The HDDs have similar results but we should single out the WD4000KS again. This HDD performs superbly here.

Small files again, and the HDDs from Seagate, Maxtor and Hitachi again have problems.

The file-sets will be read now.

We’ve got some variation at last. It is the Hitachi drives that seem to be the best here while the WD4000KS has the worst speed. Well, it’s just impossible to be always the best…

It is the physical properties of the HDD that are crucial when processing large files, so there is no wonder the HDDs from Seagate and Maxtor are on top here.

The drives have similar speeds on small files. Note the good results of the HDDs from Hitachi and Maxtor and yet another discomfiture of the WD4000KS.

Now the file-sets are going to be copied.

There’s no clear winner here, although the hitherto unremarkable WD4000YR is the fastest in two out of three arrays. Well, each of the three drives from WD did well in this test.

The Hitachi failed in RAID10, and the drives from Maxtor and Seagate in RAID5.

The WD4000KS is brilliant in RAID10, but far from that in RAID5. The Hitachi leaves the others behind when working in RAID0 – quite to my surprise.

The WD4000KS copes best with the task of copying small files, but the Hitachi is following it very closely.

I will publish the diagrams for FAT32 without any comments since they are hardly different from those you’ve just seen.


Now it’s time to draw some practical conclusion from today’s tests.

First, “professional” hard disk drives haven’t showed any great advantages over “desktop” HDDs in terms of performance. I didn’t benchmark the drives under vibration, but anyway…

Second, you could see the “benefits” from the support for NCQ technology. I mean in tests that emulate server-like operation modes of the disk subsystem.

Third, despite the declared support for NCQ by the drives from Seagate and Maxtor, it was the drive that doesn’t officially support it that really utilized that technology to a positive effect. The NCQ support in Seagate’s and Maxtor’s drives is a topic for a separate investigation.

Fourth, the Hitachi drive benefited from having five platters in one test only: when I deliberately narrowed the test’s operating zone. Otherwise, this HDD would lose either because or its low areal density or because or its small cache buffer. So, I wish I could test the T7K500 model, too.

Fifth, I recommend you to take notice of the WD drives. The company has made a real breakthrough with its xxxxKS models!

To be continued…