by Nikita Nikolaichev
02/18/2004 | 11:37 PM
Well, our life is so rich in all sorts of events that the time flies by much faster than we want it to. Having cast a glance at the calendar I was terrified to discover that 9 months have already passed since the first 10,000rpm SerialATA HDD was announced. Yes, three quarters of a year ago Western Digital released a new class of hard disk drives to the market. Featuring the same spindle rotation speed as server SCSI drives, 10,000rpm, the new WD solutions supported a pretty popular SATA interface. Of course, ATA interface was even more popular at that time, but Western Digital positioned its hard disk drive as a solution for high-performance workstations and low-end servers. These systems require HotSwap feature very insistently, so ATA interface didn’t suit there.
As for SerialATA interface, it has just settled down in our systems and not only in the form of a PCI add-on card, but also in the form of integrated chipset controllers. We do not see any grave performance differences between the SATA and ATA storage subsystems, but a combination of Raptor drives and SerialATA interface provided a truly outstanding result. Therefore, many companies which are trying top stress the highest quality of their workstations, use RAID arrays built of Raptor drives. So, we could come across WD’s raptors in almost every computer edition. I believe that most users already associate Raptor with high speed, which is quite natural in the current state of things.
But these are all external proofs of WD’s success. And what does the situation look like from the financial prospective? Hasn’t this risky experiment had a negative effect on the company’s revenues? The things here are flourishing too: the sales volumes and revenues keep growing :)
So, I have every reason to state that the success of Raptor drives strengthened WD’s reputation immensely. But they have to maintain their innovative reputation that is why new products should be released more or less regularly. This new product is the second generation Raptor from Western Digital – Raptor 74GB.
Before we start talking about the advantages of the new drives let’s recall what we were unhappy with by the old Raptor:
And now please meet our today’s hero: Western Digital WD740GD:
Compared with the first-generation Raptor drive, the new fellow features firmer and more massive stiffening ribs. But unfortunately, we again see the notorious Marvell controller on the electronics PCB :(
The new HDD features 74GB storage capacity, which is twice as large as the old Raptor drive had. This storage capacity was achieved due to the use of two platters instead of one. We even weighed the drives to make sure that our suppositions are correct:
WD360GD on the left and WD740GD on the right
As you see, the first HDD is a little bit heavier. Just one platter heavier… You can also clearly see from the photo below that WD740GD really does have two platters:
However, the differences between the new WD740GD and the old WD360GD are much more significant than just a different number of platters. The new Raptor features platters with higher bit density and hence boasts higher read speed from the platter. The track seek time also reduced to 4.5msec against 5.2msec by the old WD360GD. The 0.7msec seek time improvement is a tremendous technological breakthrough, which puts WD740DG onto one level with the most advanced SCSI HDDs with 10,000rpm spindle rotation speed.
The new Raptor also boasts one very interesting technology aka Rotary Accelerometer Feed Forward (RAFF). The idea behind this technology implies that the HDD watches the case vibrations generated during the read/write heads repositioning. This is done with the help of special accelerometers (sensors measuring the acceleration speed). Special algorithms calculate the heads trajectory so that it takes them the minimal amount of time to find the desired track without any delays. The same technology also works during linear reading: the HDD controller keeps the head on a track responding to the case vibrations with the preliminarily calculated actuator movements.
One more interesting technology from WD, which is not directly connected with Raptor drives, but which is worth mentioning here In the very first article devoted to a SerialATA hard disk drive called Seagate Barracuda Serial ATA V Hard Disk Drive Review I mentioned that the SATA cable is not locked tightly enough in the corresponding connector of the drive. If the cable hanged loosely or was slightly pushed aside, then the plug jumped off the connector and… I believe you can imagine what happened next. What a great data losses could happen if the connection between the HDD and the controller would break during the data transfer! I am afraid even to think of it, not to mention going through it on my own…
Only if the HDD and cable connectors get bigger the contact could get somewhat more secure. And if this is not possible, then all we have to do is to increase the contact area around the connector. This exactly what WD engineers did in their new Raptor drives. The new solution is called SecureConnect!
The new SecureConnect cable looks as follows:
And this is what the HDD connector look like:
The red arrows mark special holes in the HDD case where the connector hooks should go once you plug the cable in. Note that the holes are equipped with special metal pads from the inside for higher contact security.
Due to two “rails” and large contact surface of the new plug, the connection is highly secure, which is exactly what the new SecureConnect has been aimed at. I wish I had more than two cables like that…
And now let’s pass over to the discussion of one very interesting and mysterious feature of the new WD740GD drive.
Western Digital announced that second generation Raptor drives acquired Ultra/150 Command Queuing (Ultra/150 CQ) technology. Is the so long-awaited revolution? In fact, the situation is not as simple as it might seem at first glance.
I have every right to state that WD740GD drive doesn’t process the commands with NCQ, i.e. it is not the native processing of the requests queue, which was officially announced in Serial ATA II: Extensions to Serial ATA 1.0a, revision 1.1.
If you remember, the PCB of our WD740GD still has a Marvell 88i8030 chip, which means that WD740GD is an ATA drive, which can be connected to the SATA controllers because it is equipped with an ATA-to-SATA bridge.
ATA protocol acquired support of requests queue processing a while ago, when they introduced ATA/ATAPI-4 specifications. But there was only one HDD manufacturer who implemented Command Queuing in the drives: IBM (see our article called The Last IBM Drive: Deskstar 180GXP HDD Details for details). This way Command Queuing in an ATA drive is possible not only theoretically, but also practically, and the hard disk drive manufacturers can definitely put it into life if they desire to.
So, assume that the new Raptor supports ATA Queuing. However, in this case a question rises: does Marvell chip understand a set of commands implementing ATA Queuing?
Let’s check Marvell’s site. This is what we find in the specifications for the chip (click here to go to Marvell’s page):
Supports ATA command queuing
Well, this is exactly what I was striving at. But wait, and who said that CQ should only be supported by the drive? If the controller doesn’t support queue processing then it doesn’t matter at all if the HDD supports it or not. Tagged commands will never come to the drive in this case…
Assume that the SATA controller we are using supports command queuing (since SATA 1.0/1.0a specification includes this support). Also we have a PATA drive on the other end of the cable connected via a special converter. In this case the SATA controller should work as a compatible device emulating the registers of a PATA controller. As a result, only initial PATA queuing can work in this case.
Now let’s check if the diagnostics software will recognize CQ support by our Raptor drive. We will start with a very new and not very widely spread FC-IOMark utility (if you can think of a better name for it, your suggestions will be most welcome :). We haven’t yet finished working on it, but it already can disclose a lot of interesting things about the hard disk drives we run it on. We are particularly interested in the following section:
WD740GD supports CQ with up to 32 requests queue depth! Well, here we can stop looking for the implemented command queuing support, as we have just found it :)
Now let’s try to figure out if the implementation of the CQ affected the WD740GD performance in any way. The easiest and the most illustrative way to estimate the CQ efficiency is to compare the performance of the drive with the enabled CQ support with the performance of the same drive with the disabled CQ support. As a result of this simple comparison we could get exactly this particular efficiency value, but we discovered some unexpected problems on the way.
It turned out much harder to disable CQ support by WD740GD than to find it!
On reading these words many of you could have recalled that most SATA controllers use SCSI mini-port as a driver. And it means that there is a SCSI Properties page, which appears in the HDD properties window. And on this page you can see “Disable tagged queuing” option. This all absolutely true: there is a page and there is an option.
Of course, we selected this option immediately. However, our experiments showed that the HDD performance doesn’t depend on the fact whether the “Disable tagged queuing” option is selected or not.
This is unbelievable! But the surprises do not end up here. It turned out that selecting this option doesn’t have any effect even for the SCSI HDDs! We tried to check how the thing works with an Adaptec 39320D controller and different hard disk drives, and suffered a complete fiasco.
Well, it looks as if we will not manage to disable CQ in the SCSI Properties page. But let’s not give in so soon. I have already had a chance to “play” a little bit with the supported queue depth on SCSI drives by editing the registry key (see our article called Ultra320 SCSI Interface: Highs and Lows. Part II for details):
HKEY_LOCAL_MACHINE\System\CurrentControlSet\Services\adpu320\Parameters\Device\
But then I simply changed the queue depth from 32 to 64 or 256. And if we set it to 1 or 2? Will this automatically disable CQ?
Let’s check this out now! We find a section of the registry where the SATA controller driver left a track (for instance if we use a Promise S150 TX2 Plus controller):
Well, we see the magic word “Tag” and the number “33”, which stands for “32+1”. So, we change this parameter, run the benchmarks anew, and… Nothing changes! The HDD performance remained the same! Unfortunately, this method didn’t work for any of the SATA controllers we had at our disposal. This parameter of the SATA drives seems to be used for some other purposes, I assume…
Maybe we could try to estimate the CQ efficiency by comparing the dependence of the HDD performance on the queue depth? For this purpose we create a simple pattern in Intel IOMeter, which contains requests for random sectors reading, and run the benchmark a few times for different queue depths varying from 1 to 256 requests with 1 request step. We run the tests for two WD drives: WD740GD and WD360GD. If the dependence graph for the HDD performance under this type of workload will be shaped differently for both drives, then we can suppose that it is the effect of the CQ. So, what do we see?
These are two almost identical curves, though one of them is a little above the other. And where are the qualitative differences? Maybe we should pay special attention only to the part of the graph for 1-32 requests queue? No problem:
But even when we take a closer look at this given part we do not notice any principal differences in the dependence of the HDD performance on the workload. We see a “flat area” typical of the IDE drives under low workload, and a slight increase in the gap between the two testing participants as the workload grows. If this small gap is the only effect made by the CQ, then I should admit that I am pretty disappointed. But the results shown by WD740GD are most likely to be growing faster because of the faster actuator used in this HDD.
Well, let’s try and play this trick with the queue depth on a SCSI drive, just to make sure. If the graph showing the dependence of the HDD performance on the requests queue depth differs for SCSI drives with enabled and disabled CQ, then… Ok, let’s first look at the graphs:
As we see, the old trick works fine on the SCSI drives. Having reduced the queue depth for tagged-requests, we prevented the SCSI drive (in our case it was Maxtor Atlas 15K) from processing these requests in the optimal way from its point of view. That is why it is not at all surprising that the HDD performance under any type of workload laid by the Intel IOMeter pattern equaled the HDD speed for queue=1.
But look how scalable is the HDD speed when we enable TQ! Even when we had only two requests the HDD sped up a lot: and we do not see any flat spots on the curve.
So, the experiment with a SCSI drive showed us the efficiency of the TCQ technology. However, we haven’t come any closer to the question: does Command Queuing work properly by WD740GD? If the graph for WD360GD HDD, which doesn’t support CQ, were similar to that of Maxtor Atlas 15K with disabled TQ support, then we could have stated with all certainty that WD740GD wins due to the implemented CQ support. However, both WD drives proved equally scalable depending on the workload and the performance difference between them is determined only by faster seek time of the new WD740GD.
Our supposition about WD740GD requiring some special SATA 1.0 controller in order to have the TQ working properly also didn’t prove true. We tested the drive with all sorts of SATA controllers but we didn’t notice any significant performance differences.
Summing up everything mentioned above I believe there remain two versions, which sound more or less reasonable:
Here I would like to stop our discussion, but I will definitely return to this interesting topic later in the upcoming articles.
Our testbed was configured as follows:
The list of benchmarks remained the same:
We will compare our today’s hero, WD740GD drive, with the previous generation Raptor drive aka WD360GD and three SCSI drives with 10,000rpm spindle rotation speed: Hitachi IC35L073UWDY10-0 and two Seagate Cheetah 10K.6 solutions (ST336607LW and ST373307LC). This way we will be able to compare two generations of HDD predators with each other and see how their performance differs from that of the SCSI drives of the corresponding storage capacity.
The tested HDDs had the following firmware versions:
Just in case I would like to stress that we tested a pre-production sample of the WD740GD drive. That is why I would regard the obtained results as preliminary.
For our tests we used the following controllers:
In WinBench package we ran the set of tests twice for each HDD: on a 32GB partition (this logical partition was created in the beginning of the drive) and on the entire disk space. All the WinBench tests were run 7 times each and then the best result was taken for further analysis.
To estimate how greatly the “locality” affects the HDD speed, we ran the WorkStation pattern in Intel IOMeter twice: for the entire HDD space and for the first 32GB.
For FC-Test the HDDs smaller than 64GB were formatted as two equal partitions, and if the HDD was bigger than 64GB, we simply created two 32GB partitions in the beginning.
The HDDs didn’t cool down between the tests.
As usual let’s start with the longest and more informative benchmark. Here we will check the drives’ ability to work with a mixed stream of reads and writes under five types of workload. All results are summed up in the following table:
Again as usual we will discuss the results of our testing participants on the diagrams. Let’s start with linear workload:
Well,. It seems to me I have already seen something like that (see our review of WD Raptor: First ATA Hard Disk Drive with 10,000rpm Speed for details). At the same time there are a few differences. Due to a radically lower average seek time, WD740GD outperformed all SCSI competitors in RandomRead mode (of course we are only talking about those SCSI drives, which took part in our test session). In those cases when we have writes among the processed requests, the advantage of the WD740GD is also evident, except the RandomWrite mode. I assume that there is no 10K SCSI drive in the market today, which could successfully compete with WD740GD in mixed modes under linear workload.
But what will happen if we increase the workload on the drive?
As the workload grows up, the consequences are inevitable: Ultra/150 CQ of WD740GD either failed or worked inefficiently. While SCSI drives felt at home under heavier workload and immediately sped up. Of course, WD740GD is no competitor to the SCSI solutions any more.
Further workload increase helped WD740GD out, which was quite an unexpected turn for us. Of course, it still yielded to Seagate solutions, but managed to outperform the Hitachi drive in case of large writes share.
So, the results of our database tests demonstrate that CQ support by WD740GD didn’t bring any significant benefits to the drive. Although excellent lazy write algorithms remain really strong weapon of all WD HDDs.
Now let’s see how WD740GD copes with sequential reading and writing. To check this out we will be sending read and write requests (with the queue depth equal to 4) with sequentially growing address. Once a minute we will change the data block size. This testing will show us the dependence of the read/write speed on the size of the requested data block.
Here are the results obtained for sequential reading:
It is evident that WD740GD is not any slower than any of the participating SCSI drives. It actually outperforms them quite a bit. :)
At the same time, all three SCSI drives worked faster with 8KB data blocks than WD740GD. This is a pretty interesting result, as WD740GD has always been beyond any competition when working with larger data blocks. The processing of large data blocks shows very clearly how big the linear read speed difference between the two Raptor generations is.
Now let’s take a look at sequential writing:
During writing WD740GD proved nearly perfect. Although it appeared a little slower than Hitachi HDD on small data blocks, it outperforms the latter once the data blocks size increases. At the same time we can also see that its write speed on 32KB and 64KB data blocks got somewhat slower.
Now it’s high time we ran some tests in server patterns. Keeping in mind the results obtained in the DataBase pattern, we do not expect much here, but still.
In FileServer pattern WD740GD is about 10% faster than the SCSI drives under linear workload. But as the workload increases, it starts falling behind them. In WebServer pattern WD740GD drive doesn’t boast any evident advantage over the SCSI competitors under linear workload: it can’t resort to its main weapon any more – the lazy write algorithms.
If we try to average out the HDDs performance under all types of workload we will get the following picture:
No doubt that WD740GD can successfully compete with SCSI drives only under low workload, and since the rating is calculated basing on the performance of our testing participants under five types of workload, WD drives do not look that attractive any more…
Now let’s test our new Raptor in WorkStation pattern. Theoretically, the situation in this pattern is favorable for WD drives as they have to work under low loads and perform a lot of writes. Anyway, let’s check out the results:
In fact, the situation is very similar to what we have just seen in server patterns. Although the access time of our WD740GD is close to that of the participating SCSI solutions, our hero is considerably faster under low workloads. But as soon as the workload reaches 4 requests, SCSI drives outpace the winner.
On the other hand, high HDD speed under high workloads doesn’t matter that much for desktop applications, because they are not typical of desktop systems.
That is why we will calculate the performance rating for WorkStation pattern in a bit different way. Unlike server patterns where we consider all workloads to be equally probable, here we have to introduce probability coefficients for large workloads so that we could reduce their weight for the final result. The weight coefficient for the HDD performance under certain workload is inversely proportional to the queue depth.
Performance = Total I/O (queue=1)/1 + Total I/O (queue=2)/2 + Total I/O (queue=4)/4 + Total I/O (queue=8)/8 + Total I/O (queue=16)/16 + Total I/O (queue=32)/32
Now let’s see what we’ve got:
Due to a significant advantage over the SCSI drives under low workloads, the new WD740GD looks very attractive overall.
However, when we run the tests within a 32GB partition, the situation gets somewhat different: the victory belongs to a 73GB HDD from Seagate. Although WD740GD is still faster than the leader under linear workload, it is unable to break ahead in the 32GB partition, where the SCSI drive receives an indisputable bonus from the TCQ.
Now we will continue our desktop performance investigation with the help of WinBench99 tests et. I believe you are pretty familiar with this test that is why no introduction is necessary. As usual we will start our discussion with FAT32 results:
Before we go into details with our discussion of Business Disk WinMark and High-End Disk WinMark, let’s take a quick look at the Disk Inspection Test:
Note that this time we also provided two results for AAT (Average Access Time): one obtained on the entire storage capacity of the drive and another one obtained during the tests on the first 32GB of the drive.
Since a multi-platter drive involved fewer cylinders to create a 32GB logical partition, they boast better random sector seek time within these 32GB. And the larger is the HDD, the bigger is the difference between these two AAT values (if the platters are similar in both cases). Now you should understand why the manufacturers are always trying to provide the model with the highest storage capacity for review: if the test is based on “locality” then…
I would like to draw your attention to the results difference of the Hitachi drive: it has the biggest gap between the two AAT values of all the testing participants. The matter is that this is a three-platter drive, while all other our 73GB solutions are dual-platter ones. And if we use only 23GB of its storage space, then the width of the working zone per platter will be smaller than by its competitors.
As we can also see the lowest AAT belongs to WD740GD with Intel ICH5 controller: 7.7ms. The remarkable thing is that the measured AAT corresponds ideally to the calculated value: 4.7ms (average seek time) + 3.0ms (rotational latency) = 7.7ms.
Now let’s compare the linear read speed of our drives:
Well, the read speed in the beginning and in the end of the raptor drive is much higher than by its predecessor and even SCSI competitors. And the correlation between the read speed in the beginning/end of the new Raptor drive is simply excellent.
Now let’s check the performance of our drives in Business Disk Winmark and High-End Disk Winmark:
Hm.. It looks as if WD740GD has finally managed to take revenge for the failure in server patterns of Intel IOMeter. It is at least 1.5 times faster in Winbench than the SCSI HDDs. And in the Business test its advantage grew up to two times!
In case of a 32GB logical partition all HDDs performed a little bit faster overall, but the situation with the participants ranking remained the same. WD740GD is ahead of all with a huge advantage.
Now let’s take a look at the performance in NTFS file system:
Well, it looks as if only absolute performance values have changed. But wait, in the Business test WD740GD is more than twice as fast as the SCSI HDDs.
Times change, but they cannot change everything. WD hard disk drives are still very powerful in Winbench99 although they also owe a lot to the SATA controller drivers, as we have just seen.
Now we are going to try our HDDs in the last benchmark intended to check how fast they write, read and copy files. This is going to be the first time we use FC-test for a massive SCSI HDDs testing. And to tell the truth, we were really surprised with the results, I should say.
The results for WD drives were obtained with a Promise S150 TX2+ controller:
This is a very illustrative diagram, isn’t it? The write speed appeared very low for SCSI drives. Could it be the driver of our Adaptec controller? Or maybe it is the peculiarity of the SCSI solutions, because all caching algorithms are set for random workload and not for streaming requests. It is especially true for lazy writing… Anyway, last time we tested a Fujitsu SCSI drive, it demonstrated normal write speed (see our article called WD Raptor: First ATA Hard Disk Drive with 10,000rpm Speed for details).
And SATA drives from Western Digital on the contrary, performed brilliantly: the average write speed for files from the ISO pattern exceeded 40MB/sec. So, it is not at all surprising that WD740GD was faster than WD360GD on all patterns.
Now let’s compare the read speed:
Suring files reading SCSI HDDs managed to regain their reputation: only in the ISO pattern WD740GD managed to outpace SCSI drives from Seagate. At the same time the Hitachi drive, which boasts much faster linear read speed than WD740GD, fell just a little bit behind the WD solution in all patterns. Could it be that SCSI drives have something to help them during reading? Especially, when the workload is a little bit higher than the linear one…
Among the two WD drives, the newcomer is a definite leader.
Now let’s check file copying:
And what happens during files copying? Some data is read from the HDD and written into a different location. In other words, every copy operation can be split into four simple commands: read the data, move the heads, write the data, move the heads.
The results of our SCSI solutions show very clearly that the data writing is their bottleneck in this entire process, while the remaining three steps are performed at almost no expense at all.
WD drives appeared somewhat faster than SCSI drives during files copying. SCSI HDDs failed mostly on the ISO pattern, while WD740GD showed its real best there.
During file copy from one partition to another, it takes more to move the heads and the overall performance gets somewhat lower. However, it hardly changes anything: SCSI drives are defeated by their SATA rivals from Western Digital.
Now let’s see what we have in FAT32:
It is interesting that the file creation speed on SCSI drives doesn’t depend on the average file size at all and stays around 11MB/sec. For WD drives, this performance is very dependent on the average file size, but even in case of the smallest files WD740GD is almost twice as fast as any of the participating SCSI solutions. Not to mention the ISO pattern…
During files reading SCSI solutions from Seagate are again ahead. WD740GD managed to retain its leadership only in ISO pattern.
During files copying in FAT32 we witnessed only one surprise: WD360GD outperformed WD740GD. Well, things like that happen sometimes… Anyway, WD740GD is faster in NTFS :)
So, Western Digital released the second generation of their SATA drives with 10,000rpm spindle rotation speed. The new HDDs are equipped with super-dense 36GB platters and fast actuator, which determines their brilliant performance in synthetic benchmarks. The CQ support implemented in the new Raptor drives unfortunately didn’t improve the performance that much. It is hard to say why this happened: it is either the absence of CQ support by the existing controllers or the low efficiency of ATA CQ. One thing is clear though: no one is going to release any new SATA 1.0 controllers. Ultra/150/CQ support in the controller drivers is a completely different thing. Although will the SATA-controller manufacturers care about it now that the new SATA drives with NCQ support are about to come out?
Of course, Western Digital will also introduce new HDDs with NCQ support. But we do not know yet if this is going to be the “new old Raptor” or a completely new generation of HDDs with NCQ support. Especially since most SCSI HDD makers usually announce new products in winter.
Speaking about our today’s hero, WD740GD, I have to state that the new Raptor is significantly faster than its predecessor and boasts larger storage capacity, which will definitely make it more popular. Moreover, our tests showed that Western Digital WD740GD is the today’s fastest desktop hard disk drive.