|
Figure 24 – Windows 2000 SCSI and IDE buffered sequential IO performance
|
| bet | 34/47 | | Sana | 30.03.2021 | | Hajmi | 1,33 Mb. | | #13789 |
Figure 24 – Windows 2000 SCSI and IDE buffered sequential IO performance. Despite having half the rotation rate, the IDE drive compares favorably to its more expensive SCSI cousin. 2KB writes show the characteristic performance penalty. The IDE drive shows slightly more than half the throughput of its SCSI counterpart for 2KB writes.
As shown in Figure 25, buffered request processor overheads are similar for both SCSI and IDE. This is expected since requests for both the SCSI and IDE drives were offloaded to hardware controllers: an Adaptec Ultra160/m SCSI controller and a 3ware 3W-5400 controller respectively. Differences between the SCSI and IDE overheads are due to differences in their respective drivers.
Due to read-ahead and lazy writes, only part of the total request overhead is being charged to our process. Instead, the process is just being charged for the buffer copy. Therefore, measurements taken with GetProcessTimes() under-report the actual number of processor cycles used. A soaker process was needed to include cycles used by the cache manager into account. That process measures the “free” time, and so can measure the entire system overhead being consumed on behalf of SQLIO. Figure 25 shows the results measured by using the soaker.
In Windows 2000, writes still show a significant penalty for small 2KB requests (because they first read the 4KB page from disk), but the drop in overhead is much more drastic. There is no significant penalty for small requests except for 2KB writes. For small 2KB and 4KB requests, buffered requests show a lower overhead than unbuffered requests due to request coalescing. The 8KB request size is the breakeven point; at request sizes greater than 8KB, unbuffered requests have a lower overhead than buffered requests. The resulting equation is approximately:
Cpu_cost = 12μs + RequestSize* .8ns + (Request_Size/64KB)*45μs.
The first term is the cost of calling the buffer manager, the second term is the cost of the memory copy, the third term is the amortized cost of the lazy writer.
|
| |
