|
Figure 33 – Win2K two-disk hardware RAID1 sequential throughput
|
| bet | 41/47 | | Sana | 30.03.2021 | | Hajmi | 1,33 Mb. | | #13789 |
Figure 33 – Win2K two-disk hardware RAID1 sequential throughput. One of the advertised features of the 3ware card is a technology called “TwinStor” which offers RAID1+0 read performance using two drives. However, in order to achieve good rest performance, we had to make deep requests at large request sizes. It is not until there was a total of 256KB of requests outstanding did TwinStor give us any benefit. On writes, slightly less than single disk performance was achieved.
The 3ware card shows good random read IO performance with its hardware RAID1. Figure 34 shows both RAID0 and RAID1 unbuffered random performance with 8KB requests. As the graph on the right indicates, the additional request scheduling being done by the 3ware controller has improved random read IO performance significantly over that of a RAID0 stripe set. At two deep, the 3ware card shows almost linear gains from the second disk. The additional performance is due to both drives containing the same data. In the case of a RAID0 stripe set, there is a 50% chance that both random IOs will be for only one of the disks with two requests outstanding. This means for the second request, half of the time one of the disks will be idle while the other is busy servicing both outstanding requests. This can be seen in the left graph as the single disk performance of 66 IO/s only reaches 99 IO/s (66 IO/s + .5 * 66 IO/s) with two requests outstanding. The RAID1 TwinStor drive however can efficiently dispatch the second IO 100% of the time to the second disk, making full use of the second set of arms. This can be seen on the right graph as two-deep requests result in 119 IO/s or close to two times the 66 IO/s performance of a single disk.
With deep queue depths, total random read throughput is higher with TwinStor than with RAID0 for a different reason: the controller reorders the requests in a way that minimizes disk arm movement on both drives. This produces a more than doubling of random IO performance on two drives compared to one.
Write performance with TwinStor was similar to that of a single disk. This is due to the cost of mirroring: writing two copies of the bits uses both drives at the same time. The additional parallelism offered by the second drive is lost. RAID0 is not subject to this constraint as it lacks data redundancy, so it is able to take advantage of the second drive.
|
| |
