Organizing Windows for Performance

Pages: 1 2 3 4 5 6 7

Benchmark Results

This first chart represents the slowest 5 configurations in the Winstone 2001 tests, and the second represents the 8 fastest. The reason I broke them down like this is that there is a very noticeable gap between these two groups. The legend provides the details of the configuration. The boot drive and the files on that drive (remember, the OS and applications always reside on the boot drive for these tests) are listed first, followed by the second drive and files. If there is no second drive, then the one listed was the only drive used, with all files residing on that drive. If a drive has an empty parenthesis (first drive only), then the OS and apps were on that drive, and the data and swap files on the second drive. The last comment is how many IDE channels were in use (i.e., in a single channel configuration the second drive is connected as the primary slave).

In the first chart there does not appear to be anything really surprising, as the IBM drive seems to provide the best performance as the boot drive. One thing to notice, however, is that connecting the second drive as either a master or slave makes virtually no difference in these results. The implication here is that the Winstone performance is not dependent upon I/O bandwidth, as suggested above.

In this second chart, we can see that the scores are much closer to each other, and significantly higher than the first group. The first thing to notice is the lowest score on the chart. In this test, the WD 21600 (PIO 4) drive was used as the boot drive, and the IBM 75GXP contained the data and swap files. What makes this specific test so interesting is that the score is actually higher than the scores in the previous table where the IBM 75GXP was the boot drive, and the WD 36400 was the data drive.

What you cannot see from these results is the boot time and application load time. In both cases, the system was significantly slower than with the other configurations (with the exception of the other WD 21600 test in the previous chart). Not only did Windows 2000 take much longer to initialize, but also the benchmark initialization itself was noticeably slower. Referring back to the caveats given earlier, this could be a significant issue for those who frequently boot or open/close their applications.

On the other hand, those who keep their systems on 24/7 and keep frequently used apps open may see better performance by having their fastest drive as the secondary rather than the primary. Looking at all of the other configurations where the IBM drive is the data drive further bolsters this argument. In fact, when looking at both charts it becomes obvious that whenever the data resides on the fastest drive, the benchmark scores are significantly higher than when configured the other way around.

The next thing to notice is that the single channel configurations are no slower than the dual channel, and that the dual IBM setup is measurably faster than any other (as long as the data is on the second drive), including the single IBM setup. This confirms that the Winstone 2001 benchmark is not limited by I/O bandwidth, but is sensitive to disk latency. In addition, the approximately 3% better performance of the dual-IBM setup compares favorably with the results in John Howland’s RAID article referenced earlier.

So, let’s take a look at the Content Creation scores and see how they compare. Once again, there is a noticeable gap between the slowest group and the fastest. In the first chart we can see exactly the same pattern as for the Winstone 2001 scores.

In the second chart, notice again how well the configuration with the PIO 4 drive does when the data resides on a very fast drive. Also notice that the dual IBM drive setup is about 4% faster than the single drive setup – once again about the same difference we saw for RAID 0 in John Howland’s setup. There isn’t much more to be said here that has not already been said with the Winstone scores, except that Content Creation seems to be even more sensitive to I/O latency than Winstone.


Pages: « Prev   1 2 3 4 5 6 7   Next »

Be the first to discuss this article!