|
Figure 38 – Unbuffered mapped throughput over Gigabit Ethernet
|
| bet | 46/47 | | Sana | 30.03.2021 | | Hajmi | 1,33 Mb. | | #13789 |
Figure 38 – Unbuffered mapped throughput over Gigabit Ethernet. CIFS read performance is very good: at 37.5MBps it reaches PCI bus saturation on the server. It is likely that with a faster PCI bus, the throughput would be greater. Unbuffered writes showed poorer performance, with peak write performance less than half peak read performance. The curves of the write throughput graph are similar to those of mapped writes in Figure 35.
Summary
To summarize:
Win2K and both SP6 and SP3 of NT4 show similar one-disk sequential I/O performance.
NT4SP6 has no performance penalty for small-unbuffered requests. NT4SP3 shows significant penalties for both unbuffered synchronous reads and unbuffered asynchronous requests.
Basic and dynamic volumes in Win2K produce similar results.
For sequential workloads, 32bit 33MHz PCI is now the major bottleneck. PCI bus saturation can be achieved with only three disks, a configuration not uncommon on workstation machines.
Random workloads limited by the PCI bus. A PCI bus can handle about 10,000 random IOs per second, approximately the capacity of 100 disks.
Single disk IDE performance is very good, especially relative to their much more expensive SCSI cousins. Good performance is easy to attain; the rules for optimizing performance on SCSI drives still hold true for IDE. Although the 3ware card we tested was hampered by its PCI bus interface and unable to take full advantage of all four disks, second generation 3ware cards show good PCI performance and no longer have this limitation.
Single disk IDE drives are the clear winner on bang for the buck. For sequential loads, they come very near the performance of SCSI drives that are more than three times as expensive. On random IO, IDE drives have a 25% better IOs per second per dollar ratio than SCSI drives.
3ware’s IDE adapter card offers a linear improvement in sequential throughput for reads and good scalability for writes up to three IDE drives. Filling the last slot with a fourth drive gives little to no improvement in sequential IO performance on our outer band measurements due to limitations in our 3ware card’s PCI bus interface. Second generation 3ware cards have no such limitation. Random IO performance is still good, as the throughputs involved are much lower so four drives don’t come anywhere near saturating the card or bus.
IDE drives show similar overhead to that of SCSI drives, unless PIO is used. Avoid PIO as it severely degrades performance while dramatically increasing processor overhead. DMA is needed for good performance.
|
| |
