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: DataBase Pattern

It’s time to run the Database pattern which will help us examine the drives’ capability to combine write and read operations. Click here for Intel IOMeter Database Table.

The sequential load when the command requests depth equals zero is the typical scenario of interaction between your applications and the hard disk drive, so it would be most useful to watch the drives under such circumstances. The diagrams show you one member of each group of drives, whose results are typical for its group.

The newly-baked record-holder, the HDS722525VLAT80 is the fastest at each read/write combination. In fact, all the Vancouvers have similar characters, like relatives often do, and only the 7K250 models with a 2MB buffer differ somewhat. They have a smaller amount of cache memory compared to the first Vancouver, but have much more aggressive read look-ahead algorithms. As the result, the size of the cache buffer cannot accumulate and optimize requests, and lazy write becomes efficient only when there are 100% write requests in the queue.

The junior 7K250 models with an 8MB cache buffer and the previous generations of the Vancouvers show similar results. It’s good when a new product bears some traces of its origin, but it’s a pity that models of different capacities are deliberately put under unequal conditions by the manufacturer. By the way, what about the scalability depending on the load? Until now, products from Hitachi (and earlier, from IBM) used to have an excellent appetite:

It’s all right at 100% reading – just like in theory. The gaps between the participants are negligibly small.

At 100% writing the gaps have become obvious, but there are still no deviations from theory: all the drives behave in the same way as the load is growing up, while the leaders and outsiders were determined by us in the previous section already.

A more curious picture can be observed in the combo mode (the averaged results for all the loads, except the two above-described extreme cases). The very first Vancouver boasts the best scalability. The 7K250 models equipped with 2 megabytes of cache memory have a graph of the same shape, but going lower. It seems like the difference between them is due to the difference in the average seek time. The Serial ATA 7K250 drive resembles its ATA counterparts of the same buffer size (8MB), while the Vancouver 2 finds itself somewhere between the past and the future.

It transpires that the 7K250 models with an 8MB buffer increase their performance under higher loads less willingly. When measuring the low-level characteristics of the devices, we noted that they used an improved read look-ahead algorithm. Besides that, we can note that their read buffer has become smaller by a third. It may be the outcome of the employed adaptive algorithms or it may be not – anyway, this has led to a certain worsening of the scalability. On the other hand, the difference is negligibly small until four requests, and bigger loads are only typical for hard-working servers.

Let’s now see how the drives pass the patterns that simulate servers and workstations.

 
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