|
Figure 9 – Windows 2000 unbuffered asynchronous reads
|
| bet | 18/47 | | Sana | 30.03.2021 | | Hajmi | 1.33 Mb. | | #13789 |
Figure 9 – Windows 2000 unbuffered asynchronous reads. Read request depth beyond a depth of one shows a significant performance hit of 66%. Writes show similar behavior. This is likely due to the drive firmware, and not Win2K. Our Win2K tests using different drives did not show a performance penalty. Unlike NT4SP3, at one deep requests, both reads and writes receive no performance hit.
wo different things are being shown in Figure 9. The first is, unlike NT4SP3, small 2KB and 4KB unbuffered asynchronous read requests in Windows 2000 no longer incur the 33% to 66% performance hit. WCE writes showed similar results. The second is the significant performance degradation seen with requests deeper than one-deep. This is likely an artifact of the drive or driver, rather than Windows 2000, as measurements taken with different drives on different SCSI adapters don’t show any such degradation in performance.
The new dynamic volumes, introduced with Win2K, have the same performance curves as basic volumes. Throughputs and overheads remain the same except for an 8% drop in throughput for 2KB unbuffered requests. Figure 10 shows CPU overhead for a dynamic volume is slightly higher than that of a basic volume due to dmio’s longer code path lengths. The fixed cost for basic volumes is 50s per read request and 56s per write request, while the marginal cost is .46s per KB read and .51s per KB written. The fixed cost for dynamic volumes is 58s per read request and 54s per write request, while the marginal cost is .43s per KB read and .50s per KB written.
|
| |
