by Nikita Nikolaichev
07/09/2008 | 10:53 AM
Watching the never-stopping progress in the performance of central processors, graphics cards and other PC components, I sometimes wonder what component is the slowest to evolve these days. I guess it is the hard disk drive. Yes, the storage capacity of hard disks is getting higher and higher but it seems like the manufacturers are trying to engage the users into the race for gigabytes of storage not because they care so much about them. Rather, it is because increasing the capacity of the HDD has proved to be easier than increasing its performance.
The HDD becomes faster with higher recording density but only in one operation mode: when the disk is accessed for sequentially placed data. Large files are an example of that. As soon as the requested data are not below the magnetic head, the HDD has to take some time to think and move the heads. It takes a dozen or more milliseconds – a whole eternity by today’s standards. It is the delays on the part of the HDD’s mechanics that are the main obstacle to boosting the performance of HDDs. In fact, nothing new has been introduced into the head movement mechanism in the last decade, so why should there be any breakthroughs in terms of performance?
Some time ago one company, which has my respects for its attempts to increase both the capacity and the performance of hard disks (e.g. it was the first to equip desktop HDDs with a large cache buffer), introduced the unconventional product called Raptor. Targeted at desktop PCs, the Raptor combined advanced server-oriented mechanics and an appropriate spindle rotation speed (10,000rpm). The downsides were the small capacity of the disk and the amount of noise it produced. On the other hand, the average time to access data was reduced considerably from about 13 milliseconds to 8 milliseconds. That was a breakthrough indeed.
No other company has followed Western Digital’s example so far (Seagate’s Cheetah NS is not targeted at desktop PCs as is indicated even by its SAS interface). The rest of the manufacturers take no marketing risks and prefer to sell gigabytes as usual.
Solid State Drives are a new fashion on the market of hard drives. The SSD is based on flash memory and has no moving parts. Reading and writing information is performed with flash memory and the process of reading/writing takes only the time necessary to read or rewrite all the cells of the data block the disk controller can work with. The SSD has a low access time because it does not delay much when switching between different cell addresses. The SSD is actually indifferent to the type of access as it performs random-address requests as quickly as requests to sequentially placed data. The SSD is yet too expensive and its capacity is yet too small to replace the traditional HDD but its speed potential is high as you can learn from our review of Samsung’s SSDs.
So what are the ways to increase the performance of hard disk drives? Notwithstanding what I’ve written above, increasing the recording density is a way, too. With higher recording density, there is a higher chance that the necessary data are not too far from the current position of the read/write heads. In other words, the HDD has a higher chance of doing a short seek and the heads won’t have to fly across the entire platter to reach the requested data. A multi-platter design produces a similar effect – there is a higher probability that the requested data are near one of the drive’s heads.
Now I will proceed to the subject of this review, to hard disk drives with a capacity of 1 terabyte. It is the highest storage capacity available today. I don’t try to present them as anything new to you. They have been around for half a year, and the sellers and buyers and, perhaps, repair workshops should have already got used to them.
So, this review comes a little bit too late but only because I wanted to collect a full selection of 1-terabyte drives in it: each manufacturer released both “desktop” and “server” versions of its 1TB drive. Alas, the server-oriented HDDs from Samsung and Western Digital are still missing in this review.
First goes the drive that was the first to appear on the market. It is the Hitachi Deskstar 7K1000.
The developer employed the time-tested (7K400 and 7K500) five-platter design. Thus, each platter of this HDD has a capacity of 200 gigabytes. This is the smallest capacity among the HDDs to be reviewed here. Well, it was more important for Hitachi to be ahead of its competitors on the market (and indeed, Hitachi was the single manufacturer of 1-terabyte drives for a few months) than to reduce the manufacturing cost of the product. Moreover, as I said above, a multi-platter design adds to performance.
You can read our detailed review of this HDD if you haven’t done that already. Next goes the server-oriented version of the 7K1000.
It looks absolutely the same as the Deskstar.
According to the specs, the server version of the HDD features protection against vibrations, which is important for HDDs working in multi-disk cages. However, I couldn’t spot additional acceleration sensors on the Ultrastar’s PCB. They are the same as on the Deskstar’s PCB. Perhaps the point of Rotational Vibration Safeguard technology is in the algorithms the HDD’s processor uses to work with the acceleration sensors’ readings. If so, the Ultrastar differs from the Deskstar on the firmware level.
Next I will show you the upstart from Samsung. Why an upstart? Because this HDD jumped up as if from nowhere. There had been no hint of its coming, but the next moment it was right here. No one had expected this product from Samsung. Although not the first company to reveal a 1-terabyte model, Samsung produced it in a three-platter design! That was impressive compared with Hitachi’s five-platter one. This technological miracle was made possible thanks to 334GB platters from Showa Denko.
With only three platters under the hood, the drive must be quick at sequential reading, economical and quiet. From Samsung’s standpoint, the model is so fast and so perfect technically that it is called F1 as a transparent allusion to Formula One.
Seagate was somewhat tardy releasing its 1-terabyte drive. There was quite a long period of time between the announcement and the actual release. Well, some products get only better from ageing.
The eleventh generation of the Barracudas has inherited its four-platter design from the tenth (Barracuda 7200.10). Alas, Seagate could not match Samsung because Showa Denko didn’t have enough platters for every developer. On the other hand, four is fewer than five, so the HDD has dense enough platters – you’ll see it shortly.
Of course, Seagate has introduced a server-oriented version of the Barracuda.11 disk calling it Barracuda ES.2. Seagate’s naming system hasn’t changed and you can easily guess that it is the second generation of the Barracudas for enterprise systems.
The next HDD should be scrutinized thoroughly because Western Digital put an emphasis on the greenness of this model. The name of the drive implies that, too. The “GP” in Caviar GP stands for Green Power. Environment-friendly manufacture of hard disk drives is great, especially as most companies manufacture their HDDs outside the country they are registered in, but Western Digital means something different here. The Caviar GP is declared to have low power consumption in comparison with other HDDs of its class. Thus, it saves power. Bad substances are thrown into the atmosphere during the production of electricity, so saving power makes the world cleaner!
According to the Western Digital website, the Caviar GP consumes 40 percent less power than competitor products. Why? Because the Caviar GP features three smart technologies: IntelliPower, IntelliSeek and IntelliPark.
According to the description, IntelliPower technology is busy finding a balance between the spindle rotation speed, the data-transfer speed, and the caching algorithms. Being wary of such technologies, I would suppose the HDD has a lower spindle speed than the competitor HDDs and this difference is made up for by high recording density and good caching algorithms. I don’t think the HDD is trying to find an optimal rotation speed (this is possible technically, but would make it much more difficult for the HDD’s electronics). Perhaps there are a few fixed spindle rotation speeds for which the HDD has predefined latency tables. This may also mean that there will appear new models in this series that will have different spindle rotation speeds.
IntelliSeek seems to resemble Seagate’s JIT technology (Just in Time). The reading head does not move to the track with the required sector at full speed and then waits for the sector to come up. Instead, the head comes to the necessary track right at the moment the necessary sector arrives. As a result, moving the heads requires less power and the hard drive becomes less noisy overall.
That’s a superb idea but Western Digital is not the inventor of it. :)
The point of IntelliPark technology is to park the heads when the HDD is in idle mode. Yes, if the heads are parked, the aerodynamic resistance inside the case lowers, which should make it easier for the HDD’s motor. The power consumption is reduced as the consequence.
There is another aspect of this technology that will be discussed in the section about the power consumption tests.
The following table lists the firmware versions of the tested HDDs:
The test data in this review were collected using HDDs with the listed firmware. HDDs with other firmware may show different performance.
Our testing methodology hasn’t changed much. As before, we make a wide use of IOMeter and complement it with FC-Test to check the drive’s performance at processing files.
Recently we’ve added a test of benchmarking disk performance at defragmentation using Raxco Perfect Disk 8.0.
Another addition to our test program is the measurement of the power consumption of hard disk drives. You can read more about it in our article called "Hard Disk Drive Power Consumption Measurements: X-bit’s Methodology Indepth".
Besides that, the HDDs are benchmarked in Futuremark’s new PCMark Vantage suite.
Here is the complete list of the tests and applications used for this review:
Our testbed has been modified, too. We were even forced to modify it because of Samsung. We traditionally test HDDs using Promise controllers which are designed as PCI cards. And the bandwidth of the PCI bus (32 bits, 33MHz) proves to be not enough for the Samsung SpinPoint F1 drive because its linear read speed is over 110MBps (111MBps with our sample; this HDD employs adaptive formatting and the recording density for different head-surface pairs may vary in the same zone, so the linear read speed of another sample of this HDD model may be a little higher or lower).
Moreover, Samsung’s HDD proved to be highly sensitive to the interface bandwidth. We checked this out in a simple way: we performed the tests once again but installed the Promise controller into a PCI-X slot (Promise controllers support this slot at 66MHz). As a result, we had to upgrade our testbed and now it is based on an ASUS P5WDG2 WS Pro mainboard. We chose this mainboard because it had universal connectors: a free long PCI Express x16 slot (another such slot is occupied by the graphics card) and two PCI-X slots.
The Promise SATA300 TX4302 controller was installed into one of these PCI-X slots.
The testing methodology is explained in general in a separate article. I’ll give some more details about it here.
So, we test the power consumption of each hard disk drive in six modes:
The Start-up mode is meant to show you the peak power consumption of the HDD at the moment its motor is started up (as you know, the start-up current of an electromotor is significantly higher than its operating current). An important note: this test, as opposed to the others, is performed over a “cold” HDD. That is, the HDD is kept at room temperature for about 10 minutes and then connected to the testbed. This imitates the typical scenario, a cold start of the PC.
The other tests are performed over a “warmed-up” HDD. I keep the HDD running in idle mode for 10 minutes (the HDD is powered up, its platters rotating, but there are no disk accesses).
This help measure the power consumption more accurately because it depends on the temperature of the mechanical parts of the drive (the viscosity of oil in the motor and the friction force of the bearings of the heads block shaft) and the windings of the electromagnets.
The Idle mode, as explained above, imitates the situation when the PC does not access the hard disk for a while. The latter is free to spend this time in any manner. The HDD can scan its surface for potentially bad sectors, or rotate platters doing nothing, or even slip into sleep mode. The HDD has different kinds of sleep, by the way. It can keep the platters rotating but park the heads, or reduce the spindle rotation speed, or halt the spindle altogether. Of course, the deeper the sleep, the more time it takes the HDD to wake up from it and get ready for work.
The Random Read mode imitates a heavy load on the HDD. Each new read request has a new LBA address and the drive’s actuator is constantly on the move. Coupled with the typical operation mode of the motor, the load nears the maximum.
The Random Write mode is the same as the previous mode but the HDD is performing requests to write random-address data blocks. Using this load we can evaluate the effect of the deferred writing algorithms on the power consumption of a HDD. Theoretically, deferred write algorithms reorder write requests to optimize performance. Upon receiving a write request, the HDD decides if it is handy to perform the request immediately or later. As the result of the reordering, the HDD’s heads move along an optimal route instead of running wildly throughout the entire platter. Of course, the HDD spends less power when performing reordered requests.
The Sequential Read mode is designed to show the power consumption of a HDD when the latter is processing sequentially placed data (e.g. reading large files or playing video files with a high bit rate). Practice suggests that the load on the HDD electronics is quite high in this mode.
The last test mode is called Sequential Write. Like with Random Write, we’ll see if the power consumption at sequential writing and reading is any different.
So, let’s start with the Start-up mode.
These are the peak currents on the +5V and +12V power lines. These numbers cannot be used to calculate the peak power consumption of the HDDs because the maximums for each voltage do not occur at the same moment. However, these data are enough to compare the HDDs.
You can see that the HDDs from Seagate and Western Digital require quite a high current on the +12V line while the Hitachi drives consume more from the +5V line. The results are very interesting.
Theoretically, it is the Hitachi HDDs that should consume more from the +12V line because they have one platter more than the HDDs from WD and Seagate. On the other hand, the peak +12V current depends on the time the HDD takes to speed up the platters. The 1-terabyte Deskstar is declared to have a typical Power-on to Ready time of 20 seconds whereas the Caviar GP takes only 13 seconds to accelerate its platters. Of course, I should also note that the WD drive has a different spindle rotation speed as you’ll learn shortly.
The Samsung is the leader in terms of power consumption at start-up in both currents. Well, it is logical that a HDD with fewer platters consumes less power. But that’s only the start-up time. Let’s see what we have in other operation modes.
The Idle mode goes next. As noted above, we give each HDD some time (about 10 minutes) to warm up after measuring the peak currents at the start-up moment. It is only then that we perform our next measurements.
This time the results are all understandable except for the suspiciously low consumption of the WD drive: the HDDs are ranked up according to the weight of the platter pack. You can note the similar consumption of the HDDs from Hitachi and Seagate and the somewhat better results of the Samsung.
The HDD from Western Digital requires one explanation. Having analyzed the results, I decided to listen to the HDD more attentively at the moment it switched from load to idle mode. So I created a simple IOMeter pattern that was bombarding the disk with random-read requests for one minute at a request queue depth of 1. After the end of this test there were no accesses to the disk and it should have been in Idle mode. I was keeping track of the instantaneous currents by means of an oscilloscope – it is easy to track the changes visually in Idle mode due to the linear shape of the graphs.
I found a curious thing. In eight to ten seconds after the processing of the last request, the HDD clicked softly and its power consumption on the +5V line dropped suddenly! The HDD seemed to park the heads and shut them down. A regular hard disk drive usually keeps the heads alive – they are in read mode and the HDD’s controller is reading service information from the platters (to stay on the track).
As opposed to this normal behavior, the Caviar GP has power-saving priorities and sinks into a slumber at every opportunity.
Is it good or bad? Yes, it’s good from a power saving point of view. And it is not good from a performance point of view. When the HDD is in such sleep mode and receives a read/write request, it has to spend some time to wake up (unpark the heads).
Now let’s see what we have in the Random Read mode.
Hitachi’s HDDs are not quite good here. Their power consumption is not proportional to the number of platters as compared with the competitor HDDs. The Samsung is good again while the HDD from Western Digital is in the lead like in the previous test. I’ll try to explain later, using special utilities, how and why it needs so little power.
Now let’s see how many watts the HDDs can save by means of deferred writing.
So, the drives are all equal in this respect. Every model needs about 3 watts less in this test than in the Random Read mode. The HDD from Western Digital is the only one that cannot do this, but it consumes very little to start with. It would have to switch into deep sleep mode to save more power.
And now let’s check out the Sequential Read mode. The drives’ heads are not moving too much in this mode, resulting in a lower power draw compared with the Random Read mode. They still require quite a lot of power, though.
The Hitachi drives both notch 11 watts while the Samsung is slightly more economical than the Seagate models. The HDD from Western Digital is in the lead again.
The last test mode is Sequential Write.
There are no surprises here. The power consumption of nearly every drive has lowered in comparison with the Sequential Read mode. The only exception is the drive from Western Digital that has a somewhat higher power draw even. Perhaps it is due to the active operation of its electronics.
Summing up the power consumption tests, I can definitely name the winner. It is the Caviar GP from Western Digital. It has the lowest average power draw among the tested HDDs in all the test modes. The only exception is the Start-up test where the Samsung SpinPoint F1 won. The SpinPoint F1 takes second place in terms of power saving.
The power consumption tests have revealed certain oddness in the behavior of the Western Digital drive. It consumes considerably less power than its opponents in every operation mode. I’ve explained the Idle mode more or less but the other discrepancies are yet to be elucidated. I’ll try to make everything out using our internal test called IOMark.
I’ll measure the linear read speed first. We used to run the subtest from WinBench 99 for that purpose but this benchmark doesn’t support hard disks of such a large capacity. That’s sad because WinBench used to draw very pretty graphs.
So, here are the results produced by IOMark:
Hitachi Deskstar 7K1000
Hitachi Ultrastar A7K1000
The two drives from Hitachi drew almost identical graphs, indicating identical zone maps. We can note that the drives have a relatively low recording density, only 83MBps, at the beginning of the disk.
Samsung Spinpoint F1
The Samsung shows adaptive formatting as well as a record-breaking recording density of 112MBps at the beginning of the disk (it is even 118MBps on the best head/surface pair).
Seagate Barracuda 7200.11
The Barracuda 7200.11 deserves our praises for its high recording density, too. It has four instead of the Samsung’s three platters but these are very dense platters: the data-transfer rate at the beginning of the disk is over 100MBps!
Seagate Barracuda ES.2
Interestingly, the Barracuda ES.2, which is in fact a server-oriented version of the Barracuda.11, has platters with lower density at the beginning of the disk (the capacity of the platter is the same 250GB). Is it a kind of a safety margin for the disk that is supposed to work under harsh conditions?
Western Digital Caviar GP
The graph is oddly similar to the data-transfer graphs of the Hitachi HDDs. I even suspected I had made an error while saving the numbers. But on closer inspection the graphs do differ. But how can the four-platter disk from Western Digital have the same data-transfer rate as the five-platter Hitachi? They differ by a quarter in terms of platter capacity!
With this difference in platter density, the equal read speed means that the HDD with the higher platter density has a lower rotation speed. And I can measure it.
This is a screenshot of the IOMark report on the disk cache and the spindle speed. Underlined is the real spindle rotation speed of the Caviar GP.
Whatever the press releases say, the platters of this HDD rotate at a speed of 5400rpm. So, I have found one part of the secret of this economical drive. It spends less power to maintain the rotation speed.
Why did Western Digital choose this rotation speed for the GP disk? Is it because of power saving reasons? I guess not. If Western Digital had had the technical opportunity to release a 1-terabyte drive with a spindle rotation speed of 7200rpm, it would have done it. And the company did not because it didn’t have the opportunity. But all the competitors releasing 1-terabyte drives, Western Digital couldn’t help introducing its own model, trying to present its low speed as an advantage.
Next I will check out the speed at which the HDDs can communicate with their cache buffer. I will use IOMark again. IOMeter can benchmark the speed of reading from and writing into the buffer and can show the correlation between the speed and the size of the requested data block.
I should note that if I began to test the HDDs as they were right out of the boxes, the drives from Seagate and Hitachi would have been not as good as their opponents. The fact is these HDDs come with an interface bandwidth limitation. They have SATA-300 electronics but their interface speed is limited to 1.5GBps by default, i.e. to the speed of SATA-150. Why? To ensure better compatibility with old SATA controllers (integrated into mainboards or external cards). It is easy to remove this limitation on the Seagate drives: you just have to remove a jumper at the butt-end of the drive where the interface and power connectors are located. There is only one jumper there, so you can’t make a mistake. It is somewhat more difficult with Hitachi’s HDDs – you have to use the Hitachi Feature Tool for that. This is a very powerful tool, by the way. Changing the interface speed is not the only thing you can do with it.
So, in this section I will show you the results of the HDDs from Hitachi and Seagate with the interface speed limitation enabled and disabled. In the next sections the HDDs will be tested in SATA-300 mode. The pairs of disks from Hitachi and Seagate (Deskstar-Ultrastar and Barracuda 7200.11-Barracuda ES.2) have identical results, so I’ll publish the diagrams for one HDD from each pair.
The Hitachi comes first:
Hitachi Deskstar 7K1000 @ SATA150
You can see the HDD quickly reaching the interface bandwidth limit. It should do better in SATA-300 mode. There is but a small difference between the speed of reading from and writing into the cache.
Hitachi Deskstar 7K1000 @ SATA300
When in SATA-300 mode, the speed grows faster on small data blocks. The maximums of speed fall on data block sizes that are multiples of 256 (128KB).
Samsung Spinpoint F1
The Samsung drew a very nice-looking graph and surpassed the Hitachi in both burst read and burst write speeds.
Seagate Barracuda 7200.11 @SATA150
The Seagate stumbles on data blocks larger than 256 sectors. The burst write graph shows this most clear.
Seagate Barracuda 7200.11 @SATA300
With the interface bandwidth limitation removed, the Seagate pushes the burst read bar even higher, outperforming the Samsung. However, the Seagate’s dislike of large-size data blocks seems to be unrelated to the interface bandwidth.
WD Caviar GP
The HDD from Western Digital has the lowest speed of working with the cache among the tested HDDs but it is also free from problems with large data blocks (it is a problem when the burst speed is lower than the HDD’s linear speed).
So, there are two winners in this test. The Seagate boasts the highest data-transfer speed while the Samsung is fast and also indifferent to the data block size.
The next test measures the average access speed. The minimum addressable data block, i.e. one sector, is used. IOMeter is bombarding the HDD for 10 minutes with requests to read and write random-address sectors at a request queue depth of 1.
The results are surprising. The HDDs from Seagate have the best access time at reading but one of them, the Barracuda 7200.11, is just awful when doing writing!
We saw this before with Seagate’s drives when we reviewed the NL35 and Barracuda 7200.9 in our Seagate Barracuda 7200.10 review. We learned then that it was due to the HDDs doing write verification. The point of this feature is that the HDD reads the data it has just written to verify their integrity.
Seagate began to tout this feature in the Barracuda ES model and then we learned when it was enabled. According to the Seagate documentation, the write verification is enabled for Barracuda ES drives when the HDD is functioning under uncomfortable temperature conditions, when it is either too cool (below 18°C) or too hot (over 58°C).
In Russia, hot-headed revolutionaries used to be sent to Siberia to cool down a little. And we sent our Seagate drive into a thermal chamber. Setting the chamber temperature at -5°C we managed to cool the HDD to 17°C. To reach the top threshold of the write verification mechanism we set the chamber at 45°C. In both cases we could spot the verification feature getting enabled by means of Getsmart and an IOMeter pattern that was sending requests to write random-address sectors.
So, this is what we had with the Barracuda ES but not only with it. Some OEM drives from Seagate, from server and desktop series alike, verified their writing under any temperature conditions (perhaps Seagate’s OEM partners demanded more reliability).
On my part, I can claim that the HDD was under comfortable conditions (25-30°C) during my tests. The Barracuda 7200.11 can’t blame me for cruel treatment!
I can also note the good results of the Hitachi drives. They are somewhat inferior to the Seagate team at reading, but are the best at writing.
The HDD from Western Digital is the slowest in this test because its platters, as I’ve shown you above, rotate at only 5400rpm. Interestingly, if you subtract 1.36 milliseconds from the read access time of the Western Digital drive (it is the difference between the half-rotation of a 7200rpm and a 5400rpm drive), you get the read access time of the Samsung.
In the beginning of this review I mentioned possible performance benefits HDDs with more platters and heads can get. It’s time to check this out in practice.
The average access time is not the only factor comparing HDDs with different recording density (because the HDD doesn’t move its heads around the entire platter in real applications), so we have decided to introduce a new parameter that would count in the head movement speed and depend on the number of platters and on the platter density. We call it the Average Positioning Speed.
This parameter is easy to calculate: the HDD is being requested to read random-address sectors for a while but instead of averaging the time it takes to perform all the requests we average the difference between the LBA addresses of the previous and next read sector divided by the time it takes to perform this disk operation. As a result, we have the average amount of data the heads go through in each second.
Here are the results:
Hitachi’s HDDs were inferior to the Seagate in the standard test of average access time, but are faster in terms of Average Positioning Speed. We did not introduce the new parameter with the purpose of making Hitachi the winner, yet it is a noteworthy thing. Combining a good seek time with a large number of read/write heads, the Hitachi drives should benefit from short seeks – and they do!
The next graph shows the dependence between the amount of requests processed by the HDD per second and the size of the data block.
It is clear that the Seagate drives are in the lead when processing data blocks smaller than 64KB. There is a graph of the Barracuda 7200.11 on the diagram but it is mostly covered by the graph of the Barracuda ES.2. The two HDDs have the same look-ahead reading strategy.
The Hitachi drives behave similarly to each other too, but they are slower than the Seagate pair on small data blocks. The Samsung is considerably slower on small blocks but goes ahead on large ones. The HDD from Western Digital takes last place in every mode just as you could expect.
This diagram explains a lot. Here is the root of all the problems: the Barracuda 7200.11 proves to be slow only on 1- and 4-sector data blocks (0.5 and 2 kilobytes). It is not about write verification (if the HDD performed it, there would be a flat stretch at the beginning of the graph). The HDD’s processor seems to lose quite a lot of time doing deferred writing.
Interestingly, the Barracuda ES.2 is free from this problem but it is also far slower than its mate on almost every data block size. The Hitachi drives are the best in this test. The HDD from Western Digital is the slowest.
The following tests are about sequential read and write speeds. The HDD’s ability to adjust its look-ahead reading algorithms to the rate and type of requests is important here.
You can see the Samsung doesn’t feel quite good from the start. As I wrote in the Testbed and Methods section, this HDD had made us change our testbed. When I analyzed the test data, I found this HDD to be very sensitive to the bandwidth of the SATA controller. The result above was recorded when the Samsung was connected to a Promise SATA300 TX4302 controller installed into a PCI-32 slot. If the controller is plugged into a PCI-X slot (and begins to work at 66MHz rather than 33MHz), the Samsung wakes up to show its best!
Having tested all the HDDs on the new platform, I can say that some of them don’t care about the testbed modifications. The HDDs from Hitachi and Western Digital do not come near the controller’s peak bandwidth.
Another fact I found, the HDDs from Samsung and Seagate on the new and old testbed only differed in sequential speed related tests. There was no difference in tests that depended on the amount of requests performed per second.
That’s why I have not yet published the results of the HDDs on the new and old platform in this review yet. But now it’s time to do so.
The next diagram shows the performance of the two HDDs that could not reveal their potential on the old testbed. The HDDs tested on the new testbed are marked with the @66MHz suffix.
The Samsung obviously does much better with the controller seated in the faster slot. On the other hand, its speed is still far from ideal on small data blocks.
The Seagate is also faster on the faster bus, especially on 8 to 32KB data blocks. The difference between the platforms is small on large data blocks – 1.5-2MBps.
The performance drop on the new testbed at 4KB data blocks must be due to the specifics of the Promise controller’s operation in the PCI-X slot. This effect occurs in this test only.
Sequential writing goes next.
The HDD from Western Digital shows its best at writing. Despite its modest physical properties, this HDD says that performance is not all about brute force. It is far ahead of its opponents on small data blocks.
The HDDs from Seagate are in the lead on large blocks. The Samsung is not quite good again. Let’s see if these HDDs are any better on the newer testbed.
Yes, the performance improves just as with sequential reading. However, the performance gain on the faster bus is not enough for the Samsung F1 to beat the Seagate drives.
The test of the drive’s ability to reorder random-address read requests is the last one among our low-level tests. In other words, this test can show us how the HDDs can use NCQ technology. IOMeter is sending a stream of requests to read random-address data blocks. The request queue depth is steadily growing up to 32 requests.
The steeper the beginning of the graph is, the better the drive uses NCQ to improve performance.
Well, that’s a brilliant victory of Seagate’s drives. No other model can even get close to the leaders. The Hitachi drives have a flat stretch at the beginning of the graph just like PATA drives with TCQ have, but go ahead at request queue depths of over 16. The HDD from Western Digital started the test slowly but then almost reached the level of the Hitachi drives. The Samsung never really woke up in this test.
Summing up the low-level tests, I can make three interesting points:
Now I will check the drives out in more complex tests starting with multithreaded load.
Multithreaded tests seem to have become quite popular, the users discussing how good Seagate drives are at multithreading, for example. Some distinguished hardware websites have even begun to call IOMeter templates not just “Sequential Read” but “Streaming (Sequential) Read” although it is not quite clear why Streaming and why they increase the queue depth to 64 requests.
So, in this test there are one to four workers, each generating a stream of disk requests. Each worker has a dedicated address zone, so the load on the HDD is quite heavy.
Like in the sequential read and write tests, the HDDs from Samsung and Seagate were tested two times, on the old and new testbed.
When there is only one thread, the test boils down to sequential reading in 64KB blocks.
The best result belongs to the Samsung on the new testbed. The Seagate drives are somewhat slower and do not have any benefits on the new testbed. The HDD from Western Digital is the slowest here.
Now let’s add one more read thread:
The Samsung remains the leader although its speed has lowered considerably. Quite surprisingly, the HDD from Western Digital gets second place. I had not expected it to be so fast here. This proves again that the algorithms of the HDD’s processor are no less important than such parameters as recording density or cache size.
Third place goes to the Seagate working on the new platform although there is again a very small difference in performance of the Seagate drive on the two testbeds. The two drives from Hitachi are the slowest of all, the desktop version beating the server one.
The Seagate team takes top places when reading three threads. Has the developer solved the earlier problems with low performance in multithreaded mode? We’ll see at four threads. One thing you can note here is that the Samsung is a little bit slower on the new testbed.
Despite its position in the middle of the diagram the Western Digital drive is good. It loses less speed than the other drives if compared with its own performance at processing one thread.
The Seagate team is victorious again. The particular models change places in comparison with the three-thread test yet they do not differ much from each other.
The Hitachi drives are at the bottom of the diagram, accompanied by the Samsung tested on the new testbed.
Let’s see what we have at multithreaded writing now.
The two drives from Samsung and Seagate tested on the new testbed are in the lead at processing one thread. The leaders are followed by the two drives from Seagate tested on the old platform. The HDD from Western Digital is at the bottom of the diagram again.
The Samsung on the new platform is faster than when writing one thread! The same Samsung takes second place too, but it is over 18MBps slower than on the new testbed.
The HDDs from Hitachi have identical results although they differed at reading. Surprisingly, the Seagate HDDs are at the bottom of the diagram – they have been so good at multithreaded reading but fail here.
The speeds are lower when the HDDs are writing three threads, but the standings do not change. The Samsung is on top, the Seagate HDDs are at the bottom.
We’ve got the same picture at four threads. So, the HDDs from Samsung and Hitachi are the best at multithreaded writing whereas the Seagate HDDs are somewhat worse at writing than their opponents.
Comparing the absolute speeds, it turns out that the Seagate HDDs are indifferent to the type of multithreaded load. At multithreaded reading and writing they slow down by a half relative to their speed at one thread.
There are three points you can make from this test:
In this section the HDDs are tested under loads typical of the disk subsystem of a server.
As opposed to many other reviewers, we use the Database pattern to measure the performance of the HDD not at a fixed ratio of reads to writes (67% reads to 33% writes) but at 11 points, the percentage of writes changing from 0% to 100% stepping 10%.
As a result, you can choose the HDD that suits the specific ratio of reads to writes of your particular database.
I will discuss three operation modes for three request queue depths. First, the queue is 1 request deep.
The Hitachi HDDs are in the lead at low percentages of writes but the Seagate Barracuda 7200.11 overtakes them as soon as 40% writes. Interestingly, the server-oriented Barracuda ES.2 behaves more conservatively: it is somewhat slow at low percentages of writes but then improves and shows a good performance gain at high percentages of writes. It doesn’t reach the speed of the Barracuda 7200.11, though.
It is just the opposite with the Hitachi HDDs: the server version (Ultrastar) is faster than the desktop version in the Random Write mode (0% read requests).
The HDD from Western Digital is the slowest drive in this test just as you could expect.
Now the requests queue gets longer.
We’ve already seen above that the Seagate drives are good at reordering read requests. This helps them now at low percentages of writes. Take note that the Barracuda 7200.11 is ahead of the Barracuda ES.2 whereas the Hitachi drives are only competitive against the Barracuda 7200.11 in the near-30%-writes zone.
And now the queue depth is the longest.
Every HDD accelerates towards the left part of the diagram, but the Seagate HDDs are still unrivalled. The HDD from Western Digital should be noted for its being almost as fast as the Samsung, which has a higher spindle rotation speed, in nearly every mode.
So, if you are looking for a high-capacity HDD for your database, you can consider the HDDs from Seagate. They feature a good implementation of NCQ which gives them an edge against their opponents. It’s not quite clear why the desktop HDD from Seagate is faster than the server version, but that’s not my problem. :)
The results of this test suggest that Seagate’s drives are going to be superior in every other server test, but we’d better check this out in practice starting with the File-Server pattern.
In File-Server pattern there is a low percent of writes (20% to be exact), so the Seagate HDDs are going to feel quite confident in it.
Indeed, the Barracuda 7200.11 is far faster than the others whereas the Barracuda ES.2 is competing with the Hitachi drives, outperforming them under high load.
The next diagram shows detailed results for loads up to 32 requests.
You can see the Barracuda 7200.11 being slower than the Hitachi HDDs at a request queue depth of 1, but goes ahead at higher loads. The Barracuda ES.2 is competing with the Samsung at low loads.
The following diagram shows performance ratings of the HDDs calculated as the average of their results at loads of 1, 4, 16 and 64 requests.
So, the Barracuda 7200.11 delivers much higher performance than any other drive in this test whereas the Western Digital drive is the slowest here.
Next goes the Web-Server pattern. This pattern is free from write requests (this is actually not very typical of today’s web-servers that have lots of dynamic content), so the Seagate HDDs have another chance to distinguish themselves.
Rather surprisingly, this test is won by the Hitachi team. The Seagate Barracuda 7200.11 tries to compete with them at low loads (thanks to its better implementation of NCQ) but the Hitachi drives are obviously faster.
Let’s examine the range of low loads in more detail:
It is a tough struggle but the Hitachi drives are somewhat faster. We can check this out by calculating the performance ratings.
The Hitachi drives are indeed better than their opponents in this pattern.
I will use three benchmarks from Futuremark for this review because each of them has something original about itself, some special subtest. It may also be interesting for you to watch how Futuremark’s approach to benchmarking hard disk drives is evolving over time.
The point of Futuremark’s approach is in taking a certain reference HDD and running applications on it. A special program is recording a log of disk accesses. This log is referred to as a trace. This trace is then replayed on another HDD and the difference between the average response time of the reference and tested HDD is indicative of how much faster or slower the tested HDD is in comparison with the reference one. This approach is not without downsides, but it’s not the time or place to talk about them.
I will begin with one of the oldest benchmarking suites, PCMark04:
The first subtest is about booting the OS up. The HDDs from Samsung and Hitachi are on top with a considerable lead over the others. The Samsung is faster on the new testbed than on the older one. Until this test the change of the controller only affected the drive’s performance under linear loads but this test is IOps-oriented.
The Hitachi drives are superb at loading applications. The hard drive from Western Digital is also very good, competing with the Seagate drives.
The Samsung leaves no chance to the others at copying files. It is especially good on the new testbed. Once again we can see how sensitive this HDD is to the SATA controller’s bandwidth.
The Hitachi team are the best in the General Usage test. Their low response time at both random reading and writing helps cope with such a complex trace.
Futuremark developers replaced the copying test with two new tests (scanning files and writing files) in the fifth version of the suite. The other tests remained the same (the traces are different, though).
There is a new trace but old leaders. The HDDs from Hitachi and Samsung are struggling for top place.
The Application Loading test is Hitachi’s home turf. Hitachi’s drives have never lost this test as far as I remember.
The Virus Scan test is highly sensitive to the controller’s bandwidth. It means this test benchmarks the HDD’s cache, which is not quite right.
The test of writing files is sensitive to the controller’s bandwidth, too. It is won by the Samsung. As you could already see in today’s tests, this HDD prefers to have no bandwidth limitations.
Hitachi wins the General Usage test. The leader is followed by the Samsung drive tested on the new and old testbeds.
It is the first time that we use PCMark Vantage, the newest version of the benchmarking suite, for testing desktop HDDs, so I guess you should learn a little about it. The main and awful feature of PCMark Vantage is that it runs under Windows Vista only. Microsoft is rejoicing, and Futuremark is rejoicing because of Microsoft’s rejoicing, but I had to do some more preparatory work for tests under Vista.
On the other hand, the disk subtests have become much more variegated. There are no synthetic tasks like writing files. Instead, there are traces of real applications:
So, what HDD is the best in terms of the new benchmark?
The Samsung is the best at defending but is closely followed by the two HDDs from Hitachi. The pair of Seagate drives are somewhat behind the leaders while the Western Digital drive is almost as fast as the Seagate team.
The two Hitachis and the Samsung SpinPoint F1 are equals at gaming. The difference is very small between them. The Western Digital drive is at the bottom of the diagram.
The Samsung has no rivals at importing images. The pair of Hitachis is barely ahead of the Seagate team whereas the Western Digital drive is again at the bottom of the list.
The Hitachi drives win the OS Start-up test. Operating systems come and go while correct look-ahead reading algorithms stay to bring the victory to Hitachi again and again. The caching algorithms are very important as is indicated by the fact that the Western Digital drive is ahead of the Seagate HDDs here!
I have never worked in Windows Movie Maker, but I wouldn’t go near it without a Hitachi drive! I don’t know how the program loads the hard disk, but this load is obviously difficult for the drives from Seagate and Samsung. The HDD from Western Digital is good, winning third place.
The Media Center trace must be similar to the previous one in terms of load. We’ve got the same leaders: the two drives from Hitachi and one from Western Digital.
Another trace and another win for Hitachi!
The next test is Application Loading. It is easy to guess the winner.
Yes, it is yet another podium for the Hitachi team. The Seagate Barracuda 7200.11 is good, too. The Samsung lost one position in comparison with its result in the same trace from PCMark05.
Summing up the PCMark tests, I can say that when we switch to real-life applications we can have high results from those drives that have good load prediction algorithms rather than those that have record-breaking read speeds or a lowest access time or a largest cache buffer.
Hitachi has been victorious in these tests although its 1-terabyte drives are the oldest available on the market. Yes, some people don’t like them using five rather than three platters, but this doesn’t seem to be a problem.
When talking about FC-Test some people say that measuring the speed of copying files is not actually a real-life load for a hard disk. They have a point because files are not often copied within a hard disk. However, this operation is very real in terms of load type: data are being read from the disk and written to it at the same time (i.e. there are two threads, one for reading and another for writing). In fact, this is the load created by archivers (except that their threads are not symmetrical) and audio/video encoders.
What is indisputably good about FC-Test, it can create a set of files on the disk (which is the same as copying files from an external, faster, source) and read a set of files from the disk (the same as starting applications, etc). Thanks to these abilities FC-Test is very useful for testing network drives, for example.
FC-Test reacts sensitively to subtle nuances in HDDs’ firmware that are hard to reveal with synthetic benchmarks such as IOMeter. FC-Test is also good for benchmarking flash drives and HDDs with a USB interface although it’s beyond the scope of this review.
Now I will show you the results of the standard FC-Test testing procedure involving four operations over five file-sets. I won’t publish the results for the MP3 and Windows patterns to shorten this already long article. The other file-sets will show you quite clearly what HDDs are the best.
The first operation is about creating a file-set (in fact, it means writing files).
We have the same leaders at processing both large and small files. The HDDs from Samsung and Hitachi are clearly the best at writing files.
Samsung’s HDD always profits from being attached to the controller with a higher interface bandwidth. Seagate’s HDDs are surprisingly slow.
Let’s now see how the HDDs will read the file-sets.
Samsung is ahead while Seagate’s HDD does not react to the faster controller.
Reading large files almost boils down to sequential reading and the controller’s fast SATA interface is called for by both high-density models. It is only on the new platform that they nearly reach the 100MBps mark.
Seagate’s HDDs are slow again on small files while Samsung’s and Hitachi’s drives remain in the lead.
Interestingly, the copying test ranks the HDDs up in the same manner in the previous test (reading small files).
Samsung’s HDD can be proud of its own performance with the large file. No HDD has ever showed such a high copying speed.
The HDD from Western Digital copes well enough with this test while the Seagate team is slow. The identical results of the Seagate HDDs suggest that they have similar firmware.
Copying a lot of small files doesn’t change the overall picture. Yes, the Samsung drive is not so far ahead of the others, yet it is still unrivalled.
Copying the file-set into another partition doesn’t produce new results.
The numbers change but the overall picture remains the same. The Samsung is in the lead, followed by the Hitachi team. Seagate’s HDDs are at the bottom of the diagram.
The Samsung SpinPoint F1 is surely the winner of FC-Test as it combines a highest read speed with an impressive speed of writing. It has set a new record in the copying test!
I will end this test session with our homemade test of defragmentation speed. It is designed very simply: we took a HDD with one 32GB partition and installed Windows XP SP2 on it. Then we loaded the disk with music and photos, installed applications and games. We deleted some folders and installed new applications and created new files again. To cut it short, we tried to cause such a chaos that an active user of broadband Internet who doesn’t know anything about defragmenters can create on his hard disk.
We saved a per-sector copy of the disk and now copy it to the HDD we want to test. The tested HDD is connected to the mainboard’s SATA controller whose operation mode (AHCI/Standard SATA) is controlled from the mainboard’s BIOS. Next we run a FC-Test script that evokes the console version of the Perfect Disk 8.0 defragmenter and marks the time of the beginning and end of the defragmentation process.
Thus, each HDD is tested twice: with the controller’s AHCI support enabled and disabled.
Here are the results:
You can see the Hitachi drives are the only ones to benefit from the switching of the SATA controller into the AHCI mode. Considering the poor results of the Seagate drives, which have shown a good implementation of NCQ earlier, there is only one logical conclusion: NCQ has nothing to do with defragmentation. When the operation mode of the mainboard’s SATA controller changes, the disk driver changes, too. So, it is the different driver that provokes the difference in the drive’s performance in our defragmentation test.
This test is won by the Samsung, the Hitachi team taking second and third places. The Western Digital drive is good, too. Despite its lower spindle rotation speed, it competes with the Hitachi drives. This proves again how important the firmware algorithms are.
Concluding such a long rundup, I have to recall each hard disk drive with its particular highs and lows.
The two drives from Hitachi can be discussed as one because they have very similar results in every test. The Hitachi team did well in the server patterns but failed in the multithreaded tests. According to Futuremark, the Hitachi drives are ideal as hard disks for a workstation. They have been superb in every version of PCMark (thanks to their high Average Positioning Speed). In FC-Test the Hitachi team is only inferior to the Samsung SpinPoint F1, which means a lot. These HDDs are also good in our homemade test of defragmentation speed (thanks to the high APS again) but the overall impression is somewhat spoiled by the power consumption tests. 16 watts is quite a lot.
The Samsung SpinPoint F1 features high recording density, low power consumption and superb performance in the desktop-oriented tests. It has very good results in PCMark and excellent results in FC-Test and in the defragmentation test. On the other hand, it was hopeless in the server patterns. Hopefully, the server version of this drive, the SpinPoint F1 RAID, is going to be faster in the server-oriented applications. A peculiar feature of this drive is that it prefers controllers with high bandwidth. If you buy it, you should provide a fast SATA controller to reveal this drive’s potential (SATA-300 controllers integrated into mainboards will do).
The Seagate drives seem to have server roots. It looks like the developer wrote server-oriented firmware into both of them. As a result, the Barracuda 7200.11 and Barracuda ES.2 have no problems in multithreaded tests and even win at multithreaded reading. They also feel good in the Database pattern and in the File-Server and Web-Server patterns. They failed in the PCMark tests, though. The Barracudas were not brilliant in FC-Test, either, although were quite fast at reading large files. They were also slow in our defragmentation test. So, I am waiting for Seagate’s programmers to write new firmware.
Western Digital’s Caviar GP is the most ambiguous product in this review. On one hand, the manufacturer’s claim of low power consumption of this product is true. On the other hand, we can see the tradeoff just too clearly. The reduction of the spindle rotation speed to 5400rpm made it virtually impossible for this HDD to compete with the alternative products in terms of performance. Well, the Caviar GP should be given credit for challenging the other drives in some tests. It was fast in the multithreaded tests and competed with the Seagate drives in PCMark. It was quite good in FC-Test, too. In fact, this HDD can be quite a fine hard disk for a home PC because it is quiet and consumes little power. Still, I wish Western Digital added a 7200rpm model into the Caviar GP series. On my part, I promise to benchmark its performance without bias as soon as it arrives.
So, these are the highs and lows of the tested drives but the most important point about them is the price factor. The price of a 1-terabyte drive has lowered by a half since the moment they hit the market – all thanks to the competition. With four manufacturers elbowing each other aside in this market sector, we are up to more price cuts!