by Maxim Sankov
12/17/2003 | 08:24 AM
The progress is inevitable. Hard disk drives are certainly no exception to this rule. Of course, this field of computer technology has been following an extensive path lately. But let’s not forget about a well-known law of quantity transformation into quality. So, far we can single out four major ways the HDD industry can go:
If the first three ways aim at improving the consumer characteristics of the devices (the storage capacity and performance in the first place), or at least at retaining them on the same level, then the situation with Firmware optimization is not so simple at all. However, the hard disk drive manufacturers tend to resort to this tuning method more and more often today. Of course, it doesn’t make much sense to guess what the tuning like that could lead to, but you should definitely keep this “feature” of different HDD models in mind when you are making the decision about a solution you are going to use in a small server or graphics workstation, for instance. Let’s try to find out how big is the influence of these features on the hard disk drive models within a single product family (for our investigation we took four Maxtor hard disk drives with 120GB storage capacity). I wonder which of the above mentioned factors will have the most crucial influence on the overall performance of the tested solutions.
As is known there are four members in the 6Y120xx HDD family from Maxtor. They differ from one another by the size of the cache buffer and per platter density (60GB and 80GB per platter). Moreover, there are “professional” hard drives with different data density (6Y120P0) with 8MB cache buffer as well as “lite” models with 2MB cache-buffer (6Y120L0). Those of you who have been reading our site regularly, should remember that the data density per platter can be read from the HDD serial number. The first digit after the “Y” letter indicates the number of platters and consequently, their density. The fewer platters are used for the HDD of the same storage capacity, the higher is the actual data density per platter. This way, the serial number for 120GB models with 80GB platters will start with Y3…, and that of the HDD with 60GB platters – with Y4… respectively. Those of you who would like to learn more about it, should definitely consult our article called Real Maxtor DiamondMax Plus 9 HDD with 80GB Platters Reviewed!)
What other peculiarities the reviewed HDDs can boast besides higher storage capacity? The remarkable thing about these hard drives is the fact that the platters used in them will not be “cut off” according to the number of tracks, so the results of the average access time tests can be considered pretty objective this way.
Our test system was configured as follows:
We used the following benchmarking software:
All drives that support “quiet seek/normal seek” operation modes were switched to the fast mode by means of Hitachi Feature Tool. For WinBench tests, the drives were formatted in FAT32 and NTFS as a single partition with a default cluster size. We used Paragon Partition Manager for FAT32 formatting. The benchmarks were run seven times each; the maximum result was taken for further analysis. The HDDs didn’t cool down between the tests. The tests in Intel IOMeter were run in SequentialRead, SequentialWrite, DataBase, WorkStation, FileServer and WebServer patterns. If you are looking for the detailed description of these patterns, please, see our previous articles in the Storage section.
Despite the fact that HDTach has been considered not quite objective lately (and it is not for nothing that some of you may think so), I still hope it would be pretty interesting to take a look at the results in this “veteran” benchmark. Especially since the results it provides could be considered credible within a single platform tested.
What can I say here? The average (as well as the maximum) data read speed is higher on hard disk drives with higher per platter data density, while the HDDs with large cache-buffer (6Y120P0) show the minimal read speed in this test. Of course, we could blame the peculiarities of these given models or on the fact that Maxtor selects better platters for “professional” P0 solutions. However, the fact that the minimum data read speed looks like that can also indicate more efficient read ahead algorithms of P0 models, which is within firmware responsibility already. The read ahead is very likely to be working better on HDDs with 8MB cache-buffer. And the 2MB models simply have less space for prediction buffer which leads to less efficient read ahead.
We shouldn’t also disregard the possibility to disable certain firmware algorithms in L0 HDDs, again because of the too small cache-buffer. I would like to let you, our readers, discuss this topic. I would only like to point out that 6Y120P0 models haven’t shown any grave performance drops on the read-write graphs, which should be pretty typical of them, actually. Here are the pictures I am talking about:
6Y120L0-60 read graph
6Y120P0-60 read graph
These screenshots show very clearly the difference between the graphs (both: read and write graphs). They are smoother for P0 than for L0, as you may have noticed. The models with 80GB platters have the same type of graphs.
The minimal speed ratio remained the same for writing operations. As in the previous case, L0 models are almost twice as slow. During writing the models, which is supposed to be the fastest, yields to its counterpart with less dense platters in maximum speed. Though in this case the reason for the phenomenon has more to do with HDTach measuring errors. Hopefully, the next benchmark version will be free from these “special effects”.
And the leadership in the most important average value is again by the models with large per platter density.
Well, it looks as if we have now got the first surprise. And maybe this is no surprise at all? The hard disk drives with high platter density perform worse than their counterparts with 60GB platters. Although if we think a bit more indepth about the reasons for this phenomenon, we could suppose that even though the platters got denser, the rest of the HDD remained almost the same, as well as the way these platters are arranged. So the increase in the access time is a pretty logical solution.
One way or another, we can state the following: 120GB HDDs behave just the same as their predecessors tested in the previous review. Moreover, the models with lower data density are ahead in the access time tests, while those with higher data density – in sequential read and write speed tests. I should also point out that the read/write graphs for P0 models are smoother.
One more “elderly” but very useful benchmark for performance testing in everyday applications. First the table with all testing results comes:
So what do we see here? WinBench seems to prefer “professional” models with large cache-buffer. P0 models with 8MB buffer are mostly beloved by C++ test. No wonder, actually, since the reading and writing of relatively small data blocks are very often alternating during project compilation. And in Photoshop the performance difference between P0 and L0 models is no longer that big, although I expected the solutions with 80GB platters to win the leadership here. Maybe the data block size used in this patter are somewhat outdated already, which resulted into a pretty even performance of the L0 and P0 solutions with a slight advantage of the latter? We’ll see later in other benchmarks if that was the case.
I would also like to point out that the models with 80GB platters have been generally ahead of the less dense rivals here (though the advantage is not that great anyway).
Here it is: our most sensitive tool for revealing various HDD firmware peculiarities. However, this software can also be very helpful for the evaluation of other vitally important parameters of our hard drives.
The first test we are going to start with today deals with reading and writing of data blocks of various sizes. I wonder if the results in this pattern will have anything in common with what we have just seen in HDTach?
Here are the performance results for reading during Input-Output operations. This is a pretty illustrative indicator of the HDD performance under different type of workload.
And here the read speed is given in more common megabytes per second:
The models featuring large cache-buffer are indisputable leaders during small data blocks reading. However, as the data blocks grow bigger in size, the models with higher data density per platter get ahead of the others. I would like to draw your attention to the youngest 6Y120L0-60 model. As we have expected, it is a steady outsider here. Well, this is just another proof that HDTach is not completely dead. Not yet.
Now let’s pass over to the writing tests:
The advantage of the drives with large cache-buffer is really impressive, isn’t it? Well, as we see “professional” hard disk drives do not have any problems with sequential writing. However, we still have to check what happens in real applications. So far let’s have a look at the absolute write speed:
No sensation happened. The HDDs behave just the way they should. In case of small data blocks the leadership belongs to drives with large cache-buffer, while in case of large data blocks – to drives with higher platter density.
Well, now let’s pass from synthetics to more or less “real” benchmarks. We will start with database pattern under different workloads.
Why did we select load=4 and load=64? The general picture doesn’t actually depend on the workload (number of requests). I will offer you only two examples here, and if you need more info, you are always welcome to compare these results with the 80GB pervious generation drives tested in the article called Real Maxtor DiamondMax Plus 9 HDD with 80GB Platters Reviewed!
It is interesting that the whole situation changed upside down in this test. 6Y120P0-80 HDD, which has been the leader in most tests, got defeated even by 6Y120L0-60. It is pretty strange, but undeniable. Moreover, the increase in the workload doesn’t tell on the overall picture at all.
All in all, hard disk drives with 2MB cache-buffer run neck and neck. The leader in this test is 6Y120P0-60 with 8MB cache buffer. However, the performance drop suffered by the leader seems quite worrying. We should give credit to the firmware of 6Y120P0-80 for the read requests processing, however, as it comes to write requests, the problems occur. Even though the overall performance difference between the testing participants is not that big.
The situation with the graphs changes a little bit only when we get to queue depth of 256 requests, i.e. to the maximum possible number of requests in a queue. Here the leader of the DataBase pattern, 6Y120P0-60, slows down during reading, but wins back the leadership as the writes share increases. As the share of writes reaches 30%, the situation starts looking pretty familiar. Well, we could forgive the 6Y120P0-60 this dislike of the data reading under the maximum workload, since this situation is not very common for real applications. However, during intensive access of the DataBase server this performance slowdown will be quite noticeable.
Here I would like to draw your attention to a slight lag of the 6Y120P0-60 during pure reading. This data can be very useful later for further analysis.
The situation in WorkStation pattern is very similar to DataBase. So, what happened with the former leader? Is it the access time that reminds of itself here?
Now let’s check out the FileServer and WebServer patterns:
As the workload in WebServer pattern increases, 6Y120P0-80 regains its reputation, while 6Y120P0-60 starts up really greatly but then falls behind even its younger brother – 6Y120L0-60. We observe the same situation in FileServer pattern, too. Let’s draw a few parallels with the DataBase test. We saw a similar performance reduction during reading under maximum HDD workload. At the same time, we do not see anything like that by 6Y120P0-80 model featuring a cache-buffer of the same size but platters with larger data density. Maybe the firmware of 6Y120P0-60 simply turns out in a not very favorable situation with this platter density and workload type. For example, the cache-buffer string length doesn’t go well with the track size. Anyway, Firmware is exactly the one to blame for this performance drop. And in case of 6Y120P0-80 we should look into big access time value.
Well, it’s high time we summed up the performance in various Intel IOMeter patterns. During sequential reading of data blocks the HDDs run according to their performance class. However, during random data access 6Y120P0-80, the top model with 8MB cache-buffer and 80GB platters, is not always the leader, though the large cache size as well as large data density do have their effect under heavier workload. It could be one of the firmware peculiarities, although I wouldn’t exclude the possibility of a certain HDD supply. Anyway, there is one more reason explaining this behavior of the 6Y120P0-80 drive. The largest access time revealed by HDTach benchmark couldn’t remain unnoticed. It is also pretty strange that the performance of the 6Y120P0-60 model slows down under maximum workload in FileServer and WebServer patterns. Again, let’s consider Firmware responsible for that. Again, I have to repeat that the results difference is not that drastic, especially in average modes.
Now let’s have a look at the performance of our testing participants in the already traditional FC-Test developed for automatic time measurements during file copy operations in different patterns:
Here FC-Test is surprisingly similar to WinBench99. The “professional” models with large cache-buffer are ahead. Although 6Y120P0-80 managed to stand out here as well, having performed slower than 6Y120P0-60 (the solution with less dense platters) when creating multiple small files. As in Intel IOMeter, the situation improves as soon as we start working with large files and sequential access.
Reading is a true trump of the dense models! Moreover, the larger are the files, the bigger is the performance gap. HDTach did show the real state of things here, so we have every reason to trust its results for access time and read speed.
Copy near comes next. Indisputable victory of hard disk drives with 8MB cache in both: FAT32 and NTFS.
Copy far shows about the same results. It is especially noticeable when we copy large files. It looks as if we are facing true firmware optimization again. The firmware has definitely been optimized for copying of large data packs.
FC-Test benchmark results explain what we saw in Photoshop in WinBench99. since this test should include both: read and write operations of relatively large data blocks, the final results for the top P0 and L0 models appeared really close to one another.
In fact, you have just seen everything yourselves. It’s better to be “healthy and wealthy”, i.e. the combination of 80GB platters with 8MB cache-buffer ensures a pretty tangible advantage in most benchmarks. These HDDs are especially efficient for video editing or other work connected with large files processing. The second prize is won by the drive built on 60GB platters and featuring 8MB cache-buffer. The overall impression is a little bit spoilt by the performance slowdown under maximum workload in FileServer and WebServer patterns, but on the whole this is the best balanced solution of the entire family. The HDD with 80GB platters and 2MB cache-buffer is not any slower than the top model when reading large files, but yields in performance to “professionals” as it comes to writing. 6Y120L0-60 is the most “moderate” solution from the considered product family, but it also boasts a few strengths. In Intel IOMeter FileServer and WebServer patterns it performed not any worse than the rivals.
However, the average “business-user” will hardly feel the difference, I should say, so it is the price that will matter the most in this situation.
Now I will try to answer the question set in the very beginning of the review: which way of the HDD industry development is the most efficient? I believe you already know the answer. Only if all the considered options are reasonably combined together, we will get a worthy well-balanced product. A perfect example here is the 6Y120P0 solution with 80GB platters. Although there is still way for improvement in both: firmware and mechanics even by these drives.