Articles: Storage
 

Bookmark and Share

(6) 
Pages: [ 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 ]

Performance in Intel IOMeter: Average Access Time

Now let’s discuss the average read and write access time, taking two extreme cases of the Database pattern (which will be discussed later on) under a constant load of 1 request. Click here for Intel IOMeter Random Data Access Table.

The results are presented as a diagram, and we can compare these measurements to the numbers we got with the help of our own utility (see above):

It would be silly to expect an exact coincidence of the results, since our utility works with 1-sector-long data blocks while IOMeter uses 8KB blocks. Still, the order of the results is very similar, only the four single-platter models in the middle of the diagram exchanged their seats. That’s a good beginning for an early version of our benchmark, don’t you agree?

Now that we’ve ascertained the similarity between the two different measurements of the access time, we can spend some time analyzing the results. As we said above, all Deskstar 7K250 models can be divided into three groups by the access time parameter, according to the number of the data platters. The single-platter models confirm their own specifications by having the worst access time. The three-platter Vancouver 2 also joins this group – we learned in one of our previous reviews that it was the slowest in seek operations among all the Deskstar 180GXP family. Unlike in the previous generations, the three-platter models are the fastest in the Vancouver 3 series, but only due to having some tracks cut down, as we learned above. Overall, we can assume that the average seek time didn’t change since the first Vancouver.

Here we can also see the confirmation of the theory that the number of servo sectors affects the data access speed. At least, the Deskstar 180GXP, which has the lowest ratio of the amount of servo sectors to the number of tracks, is just a little bit better than the single-platter 7K250 models with a seek time slowed down by 0.3 milliseconds.

The performance when writing data to random addresses depends on the average seek time, but also on the number of requests the firmware can store in the buffer and on the efficient sorting of the requests. Again, we can divide the devices into single-, dual- and three-platter groups, with only one result falling out of the order. Well, we can’t actually give an explanation to the behavior of the 250GB SATA model.

The ratio of the average response time at read operations to that at write operations yields the efficiency of the lazy write implementation:

So, we see that the deferred write algorithm has been modernized in the Vancouver 3, just like the read look-ahead. But this is only true for the senior models on two or more platters.

The HDS722525VLAT80 model broke the record set by a Western Digital long ago. The two-platter models showed stable and high results, but the single-platter ones couldn’t even surpass the previous generations, which took their places right in the middle of the diagram. Here’s Hitachi’s policy to you – senior models have the best, and junior models are left on their own.

By the way, the Serial ATA versions are always slower than their ATA analogs. Is this a tradeoff for the reduced speed of the electronics (the Vancouver 3 has no native support of the Serial ATA interface and has to use a translator chip)? The SATA model of the highest capacity is especially slow at writing, which makes us suspect that it just has raw, experimental firmware.

 
Pages: [ 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 ]

Discussion

Comments currently: 6
Discussion started: 11/24/04 03:05:18 PM
Latest comment: 06/23/05 04:19:42 PM

View comments

Add your Comment