Figure 31 – Hardware and software RAID0 buffered throughput. In the unbuffered throughput section, we described how the 3ware’s hardware RAID0 shows better one deep, 64KB request throughput than Win2K’s software RAID0. Since FS buffering coalesces multiple requests into 64KB chunks that are issued one deep to the drive, this is reflected in the 3ware’s better buffered throughput.
Figure 32 shows the unbuffered sequential throughput of three and four Fireball IDE drives bound together using 3ware’s hardware striping. Like the two drive cases explained earlier, request depth and request size are parameters that play a large role in the throughput attained. The larger the two parameters become, the greater the throughput.
At three disks, we begin to reach the PCI bandwidth limitation of the 3ware card we tested. While reads have still scaled fairly linearly with the number of drives, unbuffered writes have plateau at 40MBps – an improvement of only 5MBps over the two drive configuration. For reads, three disks peak at 55MBps.
At four disks, the 3ware card becomes the limiting factor for sequential accesses. Writes show a 1MBps improvement over three disks, while reads show a negligible improvement. The additional fourth disk adds little to overall sequential throughput in our tests on the outer edge of the disk. However, performance at the inner diameter can still be improved substantially by the fourth disk. This can be important in video streaming applications where an important metric is the minimum streaming bandwidth across the entire disk. Although the first generation 3ware card we tested was limited by its PCI interface, second generation cards no longer have this limitation and additional drives should result in almost linear gains – up to 98MBps for reads and up to 70 MBps for writes.
|
|
|
| Figure 32 – Win2K three and four disk hardware RAID0 unbuffered sequential throughput. With three drives, reads scale fairly linearly with each additional drive, however writes plateau at 40MBps. When an additional fourth drive is added, reads show a negligible increase in performance. The 3W-5400’s PCI interface is unable to keep up with the increased drive bandwidth on the outer edge of the four disks. |
One of the advertised features of the 3ware card is a form of mirroring technology called “TwinStor” that offers RAID1+0 read performance using two drives. [Horst] Rather than reading just one drive, TwinStor spreads reads across both drives to provide RAID0 performance. As the odd-looking left graph shows, TwinStor is very much dependent upon request size and depth. It is not until there was a total of 256KB of sequential requests outstanding did TwinStor give us any benefit. TwinStor achieves its best sequential performance with deep read requests and large request sizes. On writes, slightly less than single disk performance was achieved.
The TwinStor algorithm intentionally produces this behavior. The firmware examines the request history to determine if it is mostly sequential or mostly random. Requests smaller than 256KB are treated as random to allow the second disk to independently service a second request.
|
