AMI HyperDisk100 IDE RAID Controller Review

by FastSite
01/09/2001 | 12:00 AM

IDE RAID controllers market has been developing very actively recently. These controllers are now used in all sorts of high-performance workstations and Low-End web-servers, i.e. in the most fast growing market sector. As a rule, extremely cut-throat competition is an irreplaceable part of these market spheres.<%BANNER[article]%>

In fact, we have introduced to you only one serious company, Promise, and their FastTrak and SuperTrak controllers. Today we are going to enlarge the list of your acquaintances in this field and to speak about AMI products (we don't mention HighPoint chips here, because they can't boast some features very important for fully-fledged RAID controllers).

AMI Company was founded in 1985 and is very well-known due to its high quality SCSI controllers and BIOSes, of course. However, about a year ago, on 3 February 2000, AMI announced the launching of its first IDE RAID controller called HyperDisk. And in the middle of the summer new AMI products were mentioned several times in our news, which gave us hope to see them quite soon. It was exactly at that time that there appeared an official press release from AMI and Iwill (one of the largest Taiwanese mainboard manufacturers) saying that Iwill would from then on use AMI RAID controllers on their mainboards. A new mysterious product from AMI was then known under a very beautiful and remarkable name: IDEal100. This name implied that it supported ATA/100 HDDs and was an ideal solution to be integrated into mainboards. At the same time the company launched an external IDE RAID controller - HyperDisk100.

We were looking forward to the opportunity to take a closer look at AMI controllers and to compare them with our favourites from Promise, which have won public acknowledgement and love. Finally, we managed to get our hands on a couple of pieces and today we are glad to offer you our comparative analysis of RAID standards and implementations.

Closer Look

The question number 1 that has kept troubling our mind since we first heard about AMI controllers, was the chip, which was used in them. What was it? Everything got clear as soon as we applied the so-called "lie-detector" to one of the chips:

The ordinary thermal paste served as this lie-detector. We simply spread a bit of paste over the chip surface, so that it filled all the tiny hollows on the chip. We felt relieved when we saw that there still remained only three manufacturers of IDE-controller chips: Promise, HighPoint Technologies and CMD Technology. We have already tested some controllers based on CMD chips and we have to admit that they showed very promising results. So, we sincerely hope that HyperDisk controllers will prove a worthy choice because they combine high-quality CMD chips with AMI's huge experience in writing BIOSes and drivers for RAID controllers. Although in fact, fastness is not the main thing for a good RAID controller. Sometimes it is the reliability of the disk subsystem and data protection that matter most of all.

In order to make a controller suitable for use in systems requiring high level of data protection, it (namely, its driver) should feature the following characteristics:

• Error Logging;
• Mirror Rebuilding: manual (via BIOS) and automatic (in the background mode);
• Consistency Check

AMI RAID controllers can boast all these functions.

Well, HyperDisk family currently includes two models: HyperDisk and HyperDisk100. They support UDMA66 and UDMA100 correspondingly. Here is a brief table of features for both controllers mentioned:

As you can see the controllers specs are very similar to one another and the main difference makes the supported ATA protocol. That's why we suggest taking a closer look only at the latest model of the controller, HyperDisk100, and its drivers.

The retail version of HyperDisk100 is supplied in a not very big paper box, which contains:

• HyperDisk100 controller;
• Brief user's manual;
• Floppy disk with the drivers;
• CD with the complete description and the additional utilities;
• Two 80-pin IDE cables.

All HyperDisk controllers are designed as PCI-cards (half the length of the short PCI-card) on a blue PCB.

The card is equipped with two 40-pin IDE connectors, a BIOS chip and the heart of the card - CMD chip. Besides, there are also two 2-pin connectors for the LEDs indicating the HDD status during operation.

Just below these connectors you may see a jumper enabling/disabling writing into the controller BIOS. To tell the truth we were a bit surprised to see this jumper on the card and to find nothing on AMI web-site to update the BIOS with…

BIOS and Drivers

Creating a RAID array is not always an easy task to fulfill. In other words, it's certainly not a problem to create RAID 0 or RAID 1. However, when we got AMI HyperDisk100 at our disposal and tried to create RAID 10, we failed… Here is a screenshot from the BIOS (2.4.09191628A):

As you see, there is no "Create RAID 10" command there, that's why we had to create this array manually. However, at first we couldn't think of a possible way to do it. Only when we got an Iwill mainboard with the integrated AMI controller on board, the solution was found. The BIOS used there was a newer version, which helped a lot. And then a bit later AMI updated their BIOS as well (up to 2.4.10241531A), which eliminated the difficulties once and for all.

As you may see from the screenshot above the new BIOS got a couple of hot keys to create RAID 10 arrays.

Moreover, the BIOS also acquired the possibility to enable and disable the DMA mode.

To test the performance of our controllers we used 2.4.1024 drivers. As soon as they were installed there appeared two new system devices: the controller itself and RAID 0 array:

The CD, which goes together with the controller contained two interesting programs: Spy and Microsoft Management Console Snap-In.

When they were also installed, we saw a very cute spy-man in the system tray.

If you click him with the right button of your mouse, you will get access to the following menu:

When the Spy is activated, it checks SMART of your HDDs and read/write errors from time to time. Besides, you can also run HyperDisk Console program (based on Microsoft Management Console 1.1). This program allows fine tuning of the array and of the response of the controller utilities to some possible errors:

You can also set the error polling interval there:

Now that you have a certain idea of what AMI HyperDisk controller is, let's pass over to the benchmarks.

Testbed

The testing system was configured as follows:

• Intel Pentium III 600E CPU;
• ASUS P3B-F mainboard (BIOS 1007);
• 256MB PC133 SDRAM by Hyundai;
• Fujitsu MPE3064AT HDD;
• Creative GeForce2 GTS 32MB DDR graphics card;
• Windows 98.

As far as the benchmarks used are concerned, we selected the following tools: WinBench 99 1.1 (Business WinMark and Disk Inspection Tests) and Adaptec ThreadMark 2.0. All the tests were run four times each and then the average value was taken for the diagrams and further analysis.

We carried out the tests for the following hard disk drive configurations:

• One WD273BA UDMA66 drive;
• One IBM DTLA 307015 drive;
• RAID 0 (stripe) for two IBM DTLA 307015 drives;
• RAID 1 (mirror) for two IBM DTLA 307015 drives;
• RAID 10 (stripe + mirror) for four IBM DTLA 307015 drives.

We didn't do the tests with RAID 10 for AMI HyperDisk66 and Promise FastTrak66 controllers. All HDDs were working in UDMA66 mode when the tests were run with ATA/66 controllers.

Performance with Western Digital WD273BA HDD (ATA/66)

Of course, it hardly makes any sense to buy a RAID controller for use with a single hard disk drive in a regular office system. However, a lot of today's mainboards are equipped with on-board UDMA controller chips supporting RAID functions and this tendency keeps developing quite rapidly. That's why it may be interesting to find out which controller copes with this task best of all, especially since the described situation is no longer that absurd.

The table below contains the results achieved in Disk Winbench99.

And now the results of the two major integral tests:

HyperDisk controllers proved very fast and stable in these tests. Although in case of a single hard disk drive ATA/66 controllers appeared more successful than ATA/100 ones. The performance difference between the controllers makes only 2-4%, which isn't much at all.

Also we can't disregard the fact that the CPU was very tangibly loaded when testing AMI controllers in Win-applications.

Now come the liner read graphs. Take a look at those obtained on AMI controllers: they are very beautiful, really. We have never managed to get anything like that on Promise RAID controllers. This is a definite proof that the proper drivers can guarantee uninterrupted transfer even with RAID controllers involved.

Here are the results of Adaptec Threadmark 2.0.

The results are evident. AMI controllers proved not only faster than their competitors, but also didn't cause CPU utilization increase when working. They most likely owe this success to AMI's rich experience in developing drivers for server RAID controllers.

Performance with IBM DTLA 307015 HDD (ATA/100)

In case of a slightly older UDMA66 hard disk drive from Western Digital, WD273BA, AMI controllers were a bit ahead.

Again we would like to begin with the results obtained in Disk Winbench99:

And now two integral tests for a more illustrative comparison:

Well, here Promise controllers took a revenge and managed to get the same 2-3% ahead of the competitors. Is it because of the optimization? Or inaccurate calculations? Or maybe these controllers are really better here than the rivals? :-)

And one more time Promise controllers didn't load the CPU that heavily in Disk Winbench benchmark.

Well, in Adaptec Threadmark AMI controllers aren't indisputable leaders any longer: FastTrak100 proved that Promise is a worthy rival, which still has some trumps to go. However, what price did they pay for this leadership? Over 60% CPU utilization! If we take into consideration the performance-to-CPU utilization ratio, then there will be two clearly defined groups. The first one will include AMI HyperDisk66, HyperDisk100 and FastTrak66 with the average ratio of 0.5 and the second "group" will be FastTrak100 with the ratio equal to 0.35.

So, these were the results obtained for single HDD. Now let's take a closer look at RAID arrays.

Performance with RAID Arrays

There are two major criteria used for storage subsystem evaluation: fastness and reliability (we disregard the price here). Creating a certain RAID array out of several hard disk drives allows getting a disk subsystem featuring both criteria in equal shares or a combination of them. If we are hunting for fast performance and do not care at all about the reliability of the system, then we could use RAID 0 array built of 2 or 4 HDDs. If we are looking for a highly reliable system, then RAID 1 is exactly what we need. If you want to combine both, fastness and reliability in more or less equal proportions, then you should try RAID 0+1 array. Though, note that you should have 4 hard disk drives in this case.

We decided to avoid burying you under the whole bunch of numbers that's why the diagrams below contain only those results obtained in the integral benchmarks:

And now take a look at the diagrams:

When working with fewer hard disk drives, Promise controllers surpass those from AMI, and in case of 4 hard disk drives - AMI controllers take revenge.

And here the situation is just the opposite: no matter what RAID array it is, Promise controllers are leading the competition.

Here Promise controllers again took the lead almost everywhere except the tests with RAID 0 for 2 HDDs, where HyperDisk100 managed to leave its rivals far behind. Judging by the results obtained, AMI controllers are not very successful with mirroring (RAID 1 and RAID 10), since they appear slower in these situations. It is most likely to be connected with the mirroring correctness control.

The speed reduction in case of RAID 0 array created for 4 hard disk drives is most likely to be connected with the lower maximum speed provided by AMI controllers. Note that the maximum transfer rate of ~78MB/sec is achieved in RAID 0 for both, 2 and 4 HDDs. Here are a few graphs of linear read speeds for your reference. First comes the graph for AMI HyperDisk66 tested with 4 HDDs (RAID 0):

To tell the truth, at first we thought it was one more problem of Disk Winbench99 test. However, graphs of the same shape were obtained with all the driver versions we tried, with two different controller BIOS versions. We finally believed our eyes as soon as we took another graph, for AMI HyperDisk100 tested with 4 HDDs (RAID 0):

This means that either the controller restricts the data transfer rate on the first tracks for the sake of the transfer linear character, or it happens by itself because the controller appears unable to provide higher data transfer rate.

And now take a look at the graphs obtained for Promise FastTrack100 in the same case (4 HDDs, RAID 0):

The hard disk drives as fast as IBM DTLA "overload" the controller with the data that's why it simply doesn't have enough time to transfer these volumes. The controller manages to speed up to the required level only by the time it reaches the 48th megabyte of the array, where the data density allows the controller to transfer the required 90MB/sec. Then the graph goes down only because of the data density reduction.

Besides, you can also clearly see the "speed drops" on the graph, which signal that the drivers are not balanced enough or that there are some hardware problems, which affect the controller operation. AMI controllers seem to be free from problems of the kind, although they are also very likely to be running at the maximum transfer rate they are capable of.

Moreover, the linearity of AMI controller graphs may also be connected with the calculation algorithm for the stripe-block address. As far as we understand, the curve of the kind could be obtained if there is a certain offset for each HDD during the calculation of the stripe-block address. In other words, you do the following:

and then you calculate the offset for each HDD. "0" stripe-block coincides with the physical "0" block on the first hard disk, "1" block on the second disk drive is located at "1+offset" address, etc.

This algorithm allows getting the average data density and equal data transfer rates in any part of the array. The stability of this parameter is very important when the server is operating under heavy workload.

Unfortunately, we don't know the real state of things, however, the graph looks fine and it doesn't ring a bell to us. We will do our best to try these controllers in servers in the nearest future, so have patience.

Block Size in Stripe Mode

We decided to look at the system with 2 IBM DTLA HDDs combined into one RAID 0 array. We changed the size of the stripe-block gradually and took the performance of the system in the ordinary tests. The results obtained are given below:

Here you can clearly see that making the stripe-block smaller than the cluster (which default size is 16KB for a 30GB hard disk drive) doesn't make much sense, because it will negatively tell on the controller performance. Also you shouldn't make the stripe-block size larger than the HDD cache buffer, because it also slows down the hard disk drive performance.

According to the test results, it is possible to calculate the most optimal stripe-block size, which will make the system show the maximum performance.

It is interesting that AMI controllers are most likely to work better with the hard disks with large cache buffers. However, unfortunately, IDE-drives can't boast a cache buffer greater than 2MB now. To tell the truth, this driver optimization doesn't have a good effect on the controller performance, that's why the default size of the stripe block makes 64KB.

Conclusions

During our tests we didn't face any compatibility problems and we proved that there can be a system working with several HyperDisk controllers at the same time (which appeared impossible for Promise controllers). The drivers worked stably and the linear read graphs were quite Ok, nothing extraordinary. Frankly speaking we were surprised to see that there were drivers for all the widely spread server and desktop operation systems.

The controllers tested featured very rich and diverse user friendly BIOS.

AMI HyperDisk66 and HyperDisk100 controllers showed excellent performance and proved worthy competitors to Promise products. Although they yielded a bit to their rivals, the linear read graphs easily compensated for this small drawback. In fact, when the controller is used in a workstation, the 2-percent difference will hardly matter to you, while the stable data transfer rate will be a very valuable thing especially for such operations like high-quality video capturing, for instance. And for a server solution stability and reliability appear worth their weight in gold.

It is evidently not for nothing that Iwill selected AMI controllers for their serious mainboards. Besides, the recent AMI press-release tells a bit about the company's slightly new marketing policy, according to which IDE RAID controllers are now included into MegaRAID controller family.