You should have already seen our first article devoted to the tests of High-End IDE RAID controllers: High-End IDE RAID Controllers Roundup. This time we are ready to offer you the second part. It will include three newcomers: Adaptec 2400A, 3ware 7810 and Promise SX6000. For more details on each of these controllers, please see the following reviews:

• Adaptec 2400A IDE RAID Controller Review
• 3ware 7810 IDE RAID Controller Review
• Promise SX6000 IDE RAID Controller Review

In this review we will compare all the currently available IDE RAID 5 controllers in all RAID levels supported. Before we start, I would like to point out that we didn't retest Promise SuperTrak100, Adaptec AAA-UDMA and 3ware 6400 with the new drives, which have been released since the times we reviewed these products. So, please regard them as a framework only.

Testbed and Methods

Our testbed, the one you saw on the photo above, was configured as follows:

• Intel Pentium III (Coppermine) 600MHz CPU;
• Supermicro 370DLE mainboard;
• 2 x 128MB Registered PC133 SDRAM with ECC by Micron;
• Fujitsu MPG3102AT HDD;
• Matrox Millennium 4MB graphics card;
• Windows 2000 Pro SP2.

To build arrays I used IBM DTLA 307015 HDDs with the enabled lazy write and read ahead. All controllers were tested with 64KB stripe-block. All the tests were run 4 times and then the average results were taken for the diagrams. The

For Intel IOMeter tests we used the following patterns:

The requests queue depth was changed from 1 to 256 (1, 4, 16, 32, 64, 256). Maybe some of you will consider 256 requests to be too much for the controllers of the type, however, we wanted to check what they are worth in the toughest testing conditions. With the workload like that the quality of controller firmware algorithms and its architecture on the performance is much higher than the influence of the disk subsystem implementation. I wonder how greatly could the controller performance change depending on the queue depth?

The diagram above shows that the higher is the queue depth, the more linear is the dependence of the array speed on the number of HDDs in it. And what do we actually need in our uneasy times? - Stability!

Since the possible application field for these controllers is still unknown to us, we will consider their behavior in all patterns for three types of workload: 1, 16 and 256 requests.

However, if you are interested in comparing the performance of these controllers in other modes, you can consult our previous reviews (see the links above).

Performance

RAID 0

Just for the sake of good looks I decided to include the single HDD tests into this section. Of course, I do know very well what is JBOD. However, I didn't want to make a new section in the article devoted to the controllers performance in JBOD. Please forgive me this simplification :)

So, first comes FileServer pattern:

When the workload is linear, the controller performance doesn't depend too much on the number of hard disk drives used. All of them show very slow performance growth, however, Adaptec 2400A still managed to become the first one.

Look, how greatly the graphs changed when the requests queue increased. The "first-group" controllers, such as Promise SuperTrak100, and Adaptec AAA-UDMA, refused to work any faster, while all the rest demonstrated classical linear dependence. Promise SuperTrak SX6000 controller little by little took the lead when the number of HDDs reached 5, but with the sixth HDD added, it suddenly lost all the advantage. What caused this sudden performance drop?

The results shown by 3ware controllers are so close that their graphs have coincided. Adaptec 2400A is a bit slower, and Promise SX6000 fell more noticeably behind the leader (note that even with 6 hard drives, there is no performance drop detected!). SuperTrak100 and Adaptec AAA-UDMA decided to take a short rest. :)

All in all, the three leaders performed so close to one another that is makes no sense to single out a leader.

And now comes WorkStation pattern:

So, what do we see? The same flat graphs as in FileServer pattern. And again Adaptec 2400A is ahead of all. 3ware controllers are a bit behind, while the worst result belongs to Promise SuperTrak SX6000, as it turned out even slower than its SuperTrak100 predecessor.

Again the graphs turn linear here and again 3ware controllers head the race. To be fair I should say that Adaptec 2400A doesn't lag too far behind the leaders :)

The difference between this case and FileServer pattern with the same workload lies with the Promise SX6000 controller performance. If in FileServer pattern it looked quite well, then here it kept lagging behind 3ware and Adaptec 2400A from the very beginning up to the very end. And the funniest thing is that Promise baby another time appeared absolutely unable to maintain high performance level in case of 6 HDDs.

The increase in the queue depth up to 256 requests pushed the three leaders to close their ranks even tighter. :)

And Promise SX6000 managed to fall farther behind than in the previous pattern and its lag gets greater and greater as the number of hard disk drives used in the array increases. SuperTrak100 and Adaptec AAA-UDMA controllers are still having a break.

Now let's pass over to DataBase:

In this pattern the increase in the number of the array members results into significant performance growth. It is most likely to be connected with too many write operations, which make around 33% for this pattern. The write request shouldn't necessarily be processed immediately, which means that it can be postponed till better times.

Judging by the graphs, Adaptec 2400A controller copes with this task best of all, and all the others run nearly neck and neck.

However, in case of the 16 requests workload, 3ware controllers catch up with Adaptec 2400A, while Promise SX6000 again (but very stably) falls behind. Not to mention the total Promise SX6000 controller's dislike for the sixth HDD in the array ;)

Just like in two previous patterns, 3ware controllers dash forward in case of the maximum workload (the thing that you can't see the graph for 3ware 6800 controller on the diagram doesn't mean that it is not there :)

Promise SX6000's lag behind the two 3ware brothers slowly increases from 10% in case of 1 hard drive up to 20% in case of 6 hard drives.

Hm, this is the first pattern where the controllers decided to perform differently. Adaptec 2400A seems to cope best of all with the lazy write, so that it managed to outperform its competitors pretty well. SuperTrak100 from Promise started quite alright, however, its performance changed drastically once the number of HDDs in the array reached three… maybe this problem could be solved by the BIOS or drivers update.

The advantage of Adaptec 2400A over the competitors has become a bit smaller. However, it is still very far ahead of them. The surprising thing is that Promise SuperTrak SX6000 never managed to catch up with the "first-group" controllers.

In this pattern Adaptec AAA-UDMA looks not bad at all: Adaptec seems to know the key :)

3ware controllers, which can't boast so much cache-memory, as Adaptec 2400A can. However, their performance depends on the number of array members almost linearly. Unfortunately, Promise SX6000 can't boast anything of the kind. Its lag is redoubled even more when the number of HDDs connected to it gets bigger.

Well, it's evident that the controllers performance in this pattern depends on the cache memory size and fastness.

Though 3ware controllers do not have so much cache memory as all the rest testing participants, they still manage to cope pretty successfully with large data-packs. Adaptec 2400A fell a little bit behind, though not so significantly, as Promise SX6000 or Adaptec AAA-UDMA and Promise SuperTrak100.

At the same time, Promise SX6000 controller showed highly linear dependence of its performance on the number of hard drives in the array, and seems to have potential for performance increase.

As the workload gets higher, Adaptec controller catches up with 3ware solutions.

And in case of the maximum workload, the performance of all three controllers in this pattern coincided within 1MB/sec in a 4-HDD array (only 3ware 7810 graph goes further). No doubt that 77MB/sec is a really great result for an array of 4 DTLA drives.

The RandomRead pattern is also very interesting to us. Unfortunately, not all the controllers were tested in this pattern, but there is still something worth your attention.

3ware 7810 controller takes the lead in RandomRead pattern with linear workload (while Adaptec 2400A was more successful in the pattern with at least some write requests). Promise SX6000 controller showed the third best result here.

However, with queue equal to 16 the situation changes completely! Promise SX600 is now leading (though it again failed in case of 6 HDDs). 3ware controller won the second prize, and Adaptec 2400A performed just the same way Promise SX6000 did with writes (that means that the gap between this controller and the competitors increases as there appear more HDDs in the array).

With the maximum workload the controllers again swap places: the first one here is 3ware 7810, then comes Adaptec 2400A, and then - Promise SX6000. This is how the workload may influence the controller performance…

The last pattern to consider in RAID 0 is SequentialRead. What do you think, which controller will win?

Absolutely correct! The winner is 3ware 7810, and Promise SX6000 and Adaptec 2400A are really having a tough battle now. Promise SX6000 controller was a bit faster up to a 4-HDD configuration. But as its graph shows, its performance is limited by something and when the configuration increases up to 5 or 6 HDDs, the read rate even drops.

The increase in the workload smoothens the graphs and changes the whole picture a bit. 3ware 7810, is still ahead of all, but Adaptec 2400A doesn't fall so far behind it any more. The read graph for promise SX6000 controller reached its utmost at the fourth hard disk drive, but after that the speed doesn't drop any lower.

Further workload growth doesn't change anything drastically.

Well, in the tests in RAID 0 mode we managed to prove that the second generation of IDE RAID controllers works "more correct" and predictable than the controllers from the first group tested.

I also noticed that each controller has its own pattern preferences. For instance, Adaptec 2400A controller performs excellently in patterns with lazy write, and 3ware controllers work perfectly in Sequential patterns. Promise SX6000 appeared well optimized for RandomRead pattern.

But the most important thing is that all controllers (in this or that way, of course) prove scalable when the number of HDDs in the array increases.

RAID 1 and RAID 01

You must be already bored by numerous graphs, so let's work with the tables a little bit :)

In RAID 1 3ware controllers feel quite at home (of course, you shouldn't disregard the 3ware brand TwinStor technology). But Adaptec 2400A is running just behind them.

Let's have a look at RAID 01:

Indisputable victory of 3ware 7810.

Again the victory was won by 3ware controllers, however, the Adaptec baby is close on their heels.

More writes allow Adaptec 2400A to outperform 3ware controllers, which do not have a large cache-buffer.

This is a benefit performance of Adaptec 2400A! :)

Surprising as it might seem, but Promise SuperTrak100 performed pretty well here.

In this pattern 3ware 7810 hardly has any competitors.

3ware 7810 controller is again the No.1. The second position is won by Promise SX6000. as we remember, This Promise controller coped pretty well with the random reads.

The testing participants got positioned just the same way as in the previous case.

As we have expected, 3ware controllers proved the best in RAID 1 and RAID 01 arrays. However, as the performance different shown by the new generation controllers is really small, we can assume that they are nearly equal.

RAID 5

If you are already tired of tables, let's return back to graphs :)

First comes FileServer pattern in RAID 5:

As you see, those controllers, which have Intel 960 chip for the XOR-processor are ahead.

All controllers show their scalability depending on the number of HDDs, but as we come to sixth hard drive, Promise SX6000 again starts working wonders.

With the maximum workload all controllers show similar results, but Adaptec 2400A is a little bit faster.

No, doubt, Adaptec 2400A is the leader here…

Adaptec 2400A controller is again the winner, then comes Promise SX6000 and 3ware 7810 is the third.

The increase in the workload again reconciled all the racers, so that their results are close to one another.

Now we will have a look at RAID 5 and see if Adaptec 2400A controller can apply its lazy write abilities here:

Well, it seems it can!

Promise controller looks not bad at all, having won the second position everywhere. 3ware controller didn't show any extremely high results but was very well scalable.

Adaptec seems to have pulled itself together: all the rivals are left far behind, due to excellent cache-buffer algorithms.

Only in case of higher workload the Promise SX6000 controller managed to catch up with Adaptec 2400A.

Look, how cool Promise SX6000 appeared in this pattern! Against its background all other controllers look somewhat pale. However, we'll see what happens when the workload increases:

And the increase in the workload woke up 3ware 7810 and pushed it to outpace all the competitors in a wink. However, what an awkward hump there is on the graph. All I can do is to cry out "Mea culpa!" and to start beating chest. No, it is not a measuring error, I paid special attention to this thing.

In the 5-HDD configuration we have, the 256KB data pack written onto the array falls directly onto 4 hard drives, as the stripe block is equal to 64KB in our case. The checksum is written onto the fifth HDD. In this mode there is no need in two reads, which are performed in general case when we write into RAID 5 array. So, the controller processing speed increases this way.

Why do we see this performance increase only by 3ware controller? It's hard to say… Maybe larger cache by Promise solution reduces this effect, or its firmware doesn't try to single out this "lucky combination". And all the other controllers simply do not allow connecting 5 HDDs to them as they are all four-channel ones.

Almost the same picture can be seen when the workload reaches its maximum.

Undoubtedly, the controllers here perform just like in RAID 0 array.

Even the performance drop shown by Promise SX6000 in six-drive configuration with 16 requests workload remained here.

In general, Promise SX6000 and 3ware 7810 controllers are faster than Adaptec 2400A in this pattern.

Here 3ware solution is beyond any competition, as it is perfectly optimized for Sequential requests.

The greater gets the requests queue depth, the better this controller copes with it.

The maximum read rate shown by 3ware 7810 in RAID 5 array makes 170MB/sec! Not bad, eh?

Conclusion

As I have already mentioned, the IDE RAID controller market is progressing very rapidly. There are three leading players in this market currently: Promise, Adaptec and 3ware, and all of them have already released their second generation RAID 5 controllers, which actually pushed me to carry out this second testing session.

As we have already seen, each of the new generation controllers boasts its own unique brand features. Each of them is optimized for some particular tasks.

For instance, Adaptec 2400A controller is very well optimized for writing operations, which makes it an optimal solution for data bases. However, this controller managed to cope with all our patterns very successfully.

Promise SX6000 controller seems to me to be a good choice for small web-servers, as it works quite well (in all RAID arrays) in RandomRead patterns. Its advantage over Adaptec 2400A is the support of 6 IDE RAID channels, which increases its application field. With the 6 HDDs available you can make two RAID 5 arrays or three RAID 1 arrays (system, base, log). And all of them will be fault-tolerant. There is one more combination possible: RAID 1 + RAID 01 (and even one more: RAID 1 + RAID 5 + HotSpare HDD). With a four-channel Adaptec 2400A you can create only two RAID 1 arrays or one RAID 5/RAID 01 array.

As for 3ware 7810 controller, it is almost universal. It is not only the variety of possible RAID array combinations built of the 8 HDDs connected to it that made me think so. Take a look at its performance in any type of array and with any workload, and you will se my point.

Although it has no large cache-buffer, it still performs very well in all conditions, even in the most unfavourable ones. I should also stress its brilliant performance with data sequences. One of its trumps is 64bit PCI bus support, which allows it to reach very high linear read and write rates. So, 3ware 7810 controller will be very helpful for work with large files and streaming video.