|
Figure 18 – Three and four disk striped unbuffered throughput
|
| bet | 27/47 | | Sana | 30.03.2021 | | Hajmi | 1,33 Mb. | | #13789 |
Figure 18 – Three and four disk striped unbuffered throughput. As more disks are added, additional performance gains are no longer a linear 25MBps per disk. A third disk adds an additional 20MBps while a fourth disk adds only 10MBps. The system is now becoming limited by the PCI bus limit of 87MBps and unable to take full advantage of the increased raw disk bandwidth. The 50MBps plateau for writes is likely due to the host bus adapter rather than the PCI bus.
Due to testing constraints, the four-disk configuration was tested on three dynamic volumes and one basic volume using SQLIO for striping. Despite the mixing of disks, since basic and dynamic disks have shown little to no differences in actual throughput, our comparison of striping done across dynamic and basic disks to striping on just dynamic disks is still valid.
Win2K’s dmio striping shows very little overhead. Compared with processor overhead on a single disk, striping across multiple disks costs close to nothing in processor overhead. Even with three disks running almost at PCI bus saturation, the processor shows no significant additional load. The fixed and marginal costs incurred for three striped disks was virtually identical to that of the one disk configuration.
|
|
Figure 19 – One disk and three disk striped dynamic volume unbuffered overheads. The graphs above were measured using GetProcessTimes(). They show disk striping on dynamic volumes adds no additional overhead for unbuffered requests.
|
|
| |
