|
Figure 20 – 8KB random IOs per second vs. request depth
|
| bet | 30/47 | | Sana | 30.03.2021 | | Hajmi | 1,33 Mb. | | #13789 |
Figure 20 – 8KB random IOs per second vs. request depth. Random IO performance is dependent upon the disk RPM. Compared with a 10K RPM Atlas 10K SCSI drive, the 5400 RPM Fireball IDE drive has 46% fewer RPM, a decrease similar to the decrease in IOs we measured across the two drives. The Atlas 10K however, costs more than two-and-a-half times the cost of a single IDE drive. In terms of IOs per second on a single disk, SCSI is better, but for IOs per dollar, multiple IDE drives have 57% more IOs than a single SCSI drive for the same amount of money.
Buffered random IO for IDE is similar to that of SCSI. Reads cost 22ms per MB on average, while writes cost 35ms per MB. The higher overhead for writes is again due to the cost of the extra read associated with each write request.
Unbuffered random requests cost half to a two-thirds less than their buffered counterparts. Unbuffered requests averaged slightly over 10ms per MB on both writes and reads. This cost is similar for 2, 3, and 4 disk configurations.
|
| Figure 21 – 8KB random buffered and unbuffered overhead. Buffered random writes are expensive – the overhead per MB was 270ms compared to 20ms per MB buffered reads. Reads and writes show similar unbuffered overheads at 12ms per MB. |
Multiple IDE drives provided near linear gains with each additional drive. With four disks at a request depth of one per disk, we were able to achieve over 262 IO/s for reads and 358 IO/s on writes. We measured the peak number of IOs per second possible in the system by sending random requests to each of the disks.
|
| |
