Articles: Storage

Bookmark and Share

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

Database Pattern

In the Database pattern the drive is processing a stream of requests to read and write 8KB random-address data blocks. The ratio of read to write requests is changing from 0% to 100% with a step of 10% throughout the test while the request queue depth varies from 1 to 256.

You can click the following links to view the tabled results for IOMeter: Database.

We will build diagrams for request queue depths of 1, 16 and 256.

When the request queue has the shortest depth possible, the Performance series is ahead at pure reading and at 10% writes.  At high percentages of writes the large-capacity SSDs of the Extreme series take the lead. The Extreme drives enjoy a larger advantage at reading than the Performance ones at writing, though.

As for the other products, the Nova is comparable to its predecessors at very high percentages of reads but loses to them, even to the 32GB model, at writing. The Force is always in the middle between the leaders and losers, and the Reactor is poor at both reading and writing. The latter SSD is more or less good at some mixed loads only.

When the queue grows longer, we have a completely different picture. The SandForce controller wakes up, taking the leading position at high percentages of reads and being one of the best at writing. The Samsung controller in the Performance series does not like long queue depths and can only beat the Reactor. The Indilinx-based SSDs fall into two groups again basing on their capacity, and the Nova is worse than the Extreme series at writing.

Winding up this part of our tests, we want to show you diagrams that illustrate each SSD’s performance at five different request queue depths.

The graphs are drawn to the same scale and serve as good illustrations of the SSDs’ writing performance. We are now absolutely sure that not only sequential speeds but also random writing performance of Indilinx-based SSDs depends on their storage capacity. The graphs are very stable: the firmware is flawless and predictable but there is no performance scalability at long queue depths, although these SSDs do some reordering of read requests.

The Force model’s controller can work with queue depths up to 16 requests long. It improves its performance at longer queue depths irrespective of the ratio of reads to writes. So, the controller either reorders them regardless of their type or employs some other mechanism that helps improve writing performance.

The Indilinx controller performs differently in the Nova: at longer queue depths, it loses performance sooner, i.e. at a lower percentage of writes. If this is due to its firmware, then the firmware is not good. And if it is due to the different flash memory type, the performance tradeoff is rather too large.

We know what to expect from SSDs based on the Samsung controller and the diagrams illustrate its behavior well enough. It can do read request reordering, but this leads to a sudden performance hit at mixed loads. Interestingly, the larger-capacity model delivers higher performance than its smaller-capacity cousin.

The JMF612-based Nova has a special character. It behaves not unlike the Samsung controller but has higher performance at low percentages of writes. Alas, when the share of writes is high, its speed plummets to 150-200 operations per second. On the other hand, this is as fast as HDDs can do and much faster than the previous JMF602B controller’s 2-3 operations per second.

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


Comments currently: 1
Discussion started: 07/09/10 04:22:02 AM
Latest comment: 07/09/10 04:22:02 AM

View comments

Add your Comment