and the Effect of NCQ on Their Performance
We proceed with our review of 400GB drives from WD — WD4000KD and WD4000YR from the Caviar RE2 and Caviar SE16 series. In Part One of this article we found out that these drives support NCQ. But it causes some problems with writing, which may have a negative effect on performance of these drives in applications. To find out whether it's true or not is our current task.
Performance in applications
First of all, let's see how these drives are optimized for multi-threaded operations. Advantages and disadvantages of NCQ in these drives must become apparent here. I traditionally use NBench 2.4 tests for this purpose, where 100MB files are written to disk and then read from it in several simultaneous threads — both closely and distantly located on the drive (in this case we used FAT32). The first diagram shows only mean results for reading and writing in several patterns.
Multi-threaded write performance of the WD4000KD without NCQ is worse nearly by 20% than with NCQ support. That is WD4000KD and WD4000YR drives demonstrate evidently good multi-threaded write performance with NCQ (their results are almost identical), unlike their results in ATTO Disk Benchmark. Moreover, if we have a look at the detailed diagram of multi-threaded write performance (below), we'll see that the write rate almost does not suffer from increasing the number of threads (that is simultaneously written files) and spacing them apart to different HDD sectors!
Aggressive caching of data for writing in a 16MB buffer of WD4000xx drives is confirmed by the fact that the speed of writing 100MB files in two out of six patterns (one thread and four threads spaced to different partitions) is even higher than the speed of the medium. That is as soon as the last fragment of a file being written is loaded into a large disk cache at the interface rate, the hard drive reports to the operating system that the operation is completed, though the data will be still being written to the platters for some time. But the operating system has already been notified that the write operation is completed and its resources can be freed from this task. That is the test shows the speed of executing these operations, as seen by a user working in an operating system with similar real objects in a file system (not some hypothetic limit, which could be imitated by an unpartitioned disk and IOmeter with its synthetic tests). Not all objects are cached for writing so aggressively in this test. It's a proof of strong write-back optimization of firmware. For example, a WD4000-like picture (when the speed in two patterns is obviously higher than the linear transfer rate) is demonstrated only by Samsung SP2504C, which is also notable for its aggressive caching data for writing. Among the other drives we reviewed here, only in Maxtor 7B300R0 the write speed for a single thread is noticeably faster than the speed of the medium (even though this drive is well-known for its lazy write optimization). That is along with the speed of multi-threaded operations with real files under Windows, this test actually provides another criterion for evaluating efficiency of write-back caching by a HDD buffer.
We don't have to say that NCQ helps the WD4000YR and WD4000KD drives to demonstrate best-in-class write performance in this test. They are slightly outperformed only by the excellent Samsung SP2504C. In general, they even outscore the famous Maxtor MaXLine III and Seagate NL35 400GB, which also profits much from NCQ in this write test.
However, the WD4000YR/KD drives have another important advantage over Seagate NL35 — these WD drives enjoy large performance gains from NCQ in multi-threaded reading, unlike Seagate NL35. The average performance gain from NCQ in this test for the WD4000KD reaches 52%. Not bad, eh? Thanks to this, our models under review enter the leading group in multi-threaded reading, being outperformed only by Maxtor MaXLine III with a 16MB buffer and the WD2500KS model with a single-chip controller (by the way, both of them lack NCQ support).
Now let's have a look at how the hard drives fair in old but still popular Disk WinMark 99 tests from the WinBench 99 package. We carry out these tests not only for the beginning but also for the middle part (in terms of capacity) of physical hard drives for two file systems. The diagrams display average results. It should be stressed that in this case the driver of the SiI3124 controller has optimization on the file system. It caches OS accesses and enlarges access blocks, so super-high results of this controller in Business test should not be taken to heart. :)
On the whole, NCQ plays an evil trick with both 400GB models from WD here - it significantly reduces performance in Business as well as in High-End tests compared the situation when NCQ is disabled (WD4000KD) or not supported (WD2500KS). However, WD drives cannot boast of high performance in the Business test even without NCQ, though WD4000KD (without NCQ) shares the second/third place with Samsung SP2504C in the High-End test and the WD2500KS outperforms all its competitors.
And now it's time for complex performance tests in PCMark04 and C'T H2BenchW benchmarks. They both play back prerecorded hard disk activity scripts in corresponding applications and measure the completion speed of each script. After that the results are averaged.
As WD published in its booklets, the WD4000KD drive with disabled NCQ is faster than any other drive of the same profile in the Futuremark PCMark04 disk test. When NCQ is enabled, WD4000KD performance drops by 2%. It's even lower for the WD4000YR — approximately on the level of the WD2500KS, which is actually quite good.
The same tendency, but more pronounced, is demonstrated in the similar C'T H2benchW track test, where the WD4000KD drive without NCQ is out of reach for its competitors. When its NCQ support is enabled, it loses approximately 8% of its performance, though it still outperforms the other models. Tests of this profile are favored by top WD models, it will be very difficult to remove them from this podium.
Adobe Photoshop swapping performance of the WD4000 drive also suffers from NCQ usage: the WD4000KD drive without NCQ is again a couple percents faster than with NCQ here; the WD4000YR is even slower. But the WD2500KS with a more progressive controller works performance miracles in this test, outperformed only by Maxtor and Hitachi drives and outscoring Samsung SP2504C.
In conclusion of this chapter we'd like to note that NCQ has brought no benefits to WD drives so far, except for multi-threaded performance tests — on the contrary, NCQ slows down the WD4000KD! That is you'd better disable NCQ in these WD drives for typical desktop tasks, except for specific multi-threaded tasks (for example, multimedia content).
Intel Iometer tests
We also use special patterns in Intel IOmeter to imitate hard disks operating in various applications. At first — traditional well-spread patterns, offered by Intel and Storagereview.com — DataBase, File Server, Web Server, and Workstation, which mostly characterize professional usage of hard drives in servers and serious workstations.
The obvious leaders in all the four patterns are WD4000YR and WD4000KD with NCQ support: they are practically on a par (KD is one iota better at long queues) and much faster than with disabled NCQ. By the way, the WD4000KD drive with disabled NCQ demonstrates nearly the same performance as the WD2500KS, which does not use NCQ. We have also published results of the Seagate NL35 400GB, which is clearly slower than these WD drives, irregardless of NCQ support. But this Seagate drive is noticeably better than our WD drives in linear access rates and average seek times (with NCQ).
Let's have a look at the mean results in these patterns and four queue depths (1, 4, 16, and 64). The data was averaged geometrically without weight coefficients.
Thus, in compliance with the WD booklet, 400GB hard drives from WD demonstrate the highest performance in these tasks. WD4000KD with NCQ support turns out even a tad faster than the professional WD4000YR. But if we disable NCQ in these drives, the average performance will drop by nearly 30%!!! Food for thought. Perhaps, NCQ problems in these WD drives are not as bad as we thought after the desktop tests. ;)
Tests in the Workstation pattern confirm the above said.
Now it's turn for our IOmeter patterns, which are closer to users of regular desktop PCs and near-line applications, though the random character of the disk access in these patterns is adequate to a professional profile as well.
The leader in the imitation of random reading and writing of large files is WD4000KD, but with disabled NCQ! When NCQ is enabled, WD drives lose indecently much performance. Write performance with enabled NCQ corroborates the problems, previously revealed by these drives in ATTO Disk Benchmark. Anyway, these WD drives are faster than Seagate NL35 of the same capacity. :)
The Seagate NL35 400GB is a tad better at reading small files, though both WD drives are practically on a par with it, if they use NCQ. If it's not used, their performance goes down. Small file write performance brings us back to the results of ATTO test: WD Caviar SE16 drives (WD4000KD and WD2500KS) without NCQ fair very well. But as soon as NCQ is enabled and there appears a command queue for writing (in this case — with relatively small blocks), performance of the WD4000KD and the WD4000YR drops fast. By the way, that's not the first time when we see these supposedly twin drives demonstrating noticeably different write performance with NCQ. The WD4000KD with enabled NCQ shows us that the higher is the queue depth, the more conflicts between NCQ "from WD" and the traditional write-back algorithms of firmware, which reduces the write performance.
In case of copying large files at random addresses across the entire disk, WD drives are slightly outperformed by Seagate NL35. NCQ support on the whole (like in case of reading/writing large files) has a negative effect on WD performance (by the way, the same concerns Seagate as well). Note that the graphs are practically horizontal with disabled NCQ, that is there is no evident dependence on the queue depth. But it's WD drives, which performance benefits from NCQ support in the small file copy test: copy performance jumps high with the queue depth growth. There is a reason for it. ;)
Results of all drives in these six tests, averaged by the queue depth, are published on the three following diagrams. In particular, we can see here that NCQ support in these WD drives has no positive effect on the large file write or read performance. To be more exact, its effect is negative. But on the other hand, a positive effect from NCQ in WD drives is demonstrated in reading and copying small files.
By the way, the situation with Seagate NL35 (its NCQ pros and cons) is similar and at the same time different from the WD situation.
According to the geomean of the six previous patterns (reading, writing, and copying files at random addresses) with the queue depth of 1-64, the fastest drive is Samsung SP2504C (thanks to its excellent results in copying large files). WD drives follow the leader - the WD4000KD looks a tad better without NCQ, the WD4000YR is evidently outperformed by the WD4000KD, the WD2500KS, and even one-year-old WD3200JD. But the other competitors are left far behind! :)
And finally, the NCQ situation with the WD4000 is ambiguous in the pattern of streaming read/write in large or small blocks (e.g. typical of systems for mining, storing, and broadcasting multimedia data in near-line data storage systems and servers, as well as PC or workstation operations when editing video/audio data). On one hand, performance of these drives with NCQ support improves at large blocks when the queue depth grows. But on the other hand, their performance with small blocks obviously gets lower. In case of multimedia data, we should focus on large blocks, where the NCQ effect on WD drives is positive. By the way, there is also a positive effect from NCQ in multi-threaded read/write performance tests (see above). On the whole, these WD drives are not in the lead here - they are outperformed by Samsung and Maxtor drives, though they can boast of their results versus similar models from Seagate and Hitachi.
Acoustic noise and power consumption/heat dissipation
The manufacturer claims that these drives are designed for lower operating temperatures (that is they dissipate less heat). Unfortunately, our own detailed analysis of this issue does not confirm this point: unlike less capacious WD drives of these series, which indeed dissipate much less heat during their operation than most competitors (you can have a look at this article: 6-7 Watts in idle mode and 9-10 Watts in active seek), the 400GB drives from WD are rather voracious (8 Watts for idle and 11.5 Watts for the seek mode) and hotter in operation than many competitors of the similar capacity. But, they are not the most voracious models, of course, especially in seek mode.
The same applies to the noise generated by the 400GB models from WD: according to the manufacturer, they are very quiet; but in practice, they are much less quiet than even 250-320GB WD models — both in idle and seek modes. Low-capacity WD models are indeed ones of the quietest desktop drives so far, but unfortunately, we cannot say the same about the WD4000KD and WD4000YR drives. We clearly heard humming from rotating platters and "low" chirring in active seek mode (we have tested six samples of these drives).
So if you plan on using these drives at home, get ready to put up with an increased noise level and to provide better cooling for these models (probably even use an additional fan).
So, 400GB drives of different series from Western Digital — WD4000YR Caviar RE2 and WD4000KD Caviar SE16 are very much alike. We might have thought that they are twins, but for somewhat different (in some tasks) performance (WD4000YR is outperformed) and some peculiarities of the WD4000YR that have to do with its objectives (RAID optimization, better reliability, longer warranty).
We should certainly congratulate WD with gaining another level in HDD manufacturing: 400GB capacity (a year after Hitachi and Seagate), 4-platter configuration (now WD can launch similar 500GB models without much effort; we are looking forward to such models, as we expect new performance records in our tests ;)). Besides, these two drives are the first WD models to support NCQ. However, what concerns NCQ, WD complicates things with its support for its desktop drives for a reason: according to our multiple tests (and tests, conducted by the manufacturer itself), the current models from this company noticeably lose performance with NCQ in the majority of typical desktop tasks. More detailed analysis, in this article in particular, reveals a reason for this drop — during some write operations NCQ support in WD firmware evidently conflicts with the fine-tuned and highly efficient algorithms of write-back caching in these drives, equipped with 16MB data buffers at that. By the way, we have practically no gripes with NCQ support for read operations in these WD drives. In particular, it's a reason why the 400GB WD models profit from NCQ in many professional applications (server loads and multi-threaded operations with large files). They sometimes enjoy 20-50% performance gains! These WD drives can be no worse than the leaders in typical desktop operations (they may even become the leaders here), if you disable NCQ support on the level of a host controller.
However, we do not bid farewell to these WD drives yet. Next time we'll try to test their RAID optimization compared to their direct competitors.
Alex Karabuto (firstname.lastname@example.org)
May 31, 2006
Write a comment below. No registration needed!