Revisiting the Caviar WD1000BB and the 'Cuda ATA IV

Revisiting the Caviar WD1000BB and the 'Cuda ATA IV
Introduction
Low-Level Measurements
STR Graph
WinMarks
IOMeter Performance
IPEAK RankDisk
Experimental IOMeter Indices
Conclusion

Single Page Version

Revisiting the Caviar WD1000BB and the 'Cuda ATA IV

October 8, 2001

Author: Eugene Ra

Experimental IOMeter Indices

Our File Server, Workstation, and Database IOMeter access patterns play through on all tested drives at queue depths of 1, 4, 16, 64, and 256. To simplify this intimidating benchmark, we introduced the IOMeter indices, a normalized average of results drawn from 16 I/Os, 64 I/Os, and 256 I/Os. Why normalized? Result under a load of 256 I/Os on a given drive are always higher than those drawn under 64 I/Os. Correspondingly, 64 I/O results are always greater than those of 16 I/Os. Normalization simply entails multiplying the lower load scores by a coefficient to bring their weight in line when averaging.

Running some raw trace files through IPEAK's AnalyzeTrace suggests that lower-load results are more relevant to single-user disk performance. As a result, we've normalized and averaged results from 1 I/O, 4 I/O, and 16 I/O loads under an "experimental" index. A full sort of all tested drives may be found here.

An interesting shuffle occurs when sorting today's 7200 RPM ATA drives in these new indices. IBM's Deskstar 60GXP, a drive that manages to wrap up top spots in all our deployed indices, finds itself under assault by the WD1000 series. Differences between the BB-SE and BB here remain consistent with the deployed index: identical performance in Workstation and Database patterns yet a measurable increase in File Server performance.
The Barracuda ATA IV with AAM off exhibits uniform performance improvements under all three new indices. In a notable twist, however, the 'Cuda IV has trouble keeping up with the competition in these experimental averages even with AAM disabled.
Conclusion...