Although Native Command Queuing is most definitely an elegant and intelligent way to minimize the impact of a hard drive's mechanical latency, our test results show that better performance certainly isn't guaranteed with the Barracuda 7200.7 NCQ and Promise's FastTrak TX4200. The best applications for NCQ are undoubtedly the kinds of multi-user environments that IOMeter simulates. With an increasing number of concurrent I/O requests to read or write data from different areas of a disk, command queuing's optimized request rescheduling really shines. NCQ also scales better with lower I/O loads than Western Digital's Tagged Command Queuing, whose CPU utilization is also much, much higher.
Unfortunately, IOMeter is about the only place where NCQ seems to help the Barracuda 7200.7. In fact, command queuing actually hurts the drive's performance in typical single-user applications, sometimes by a significant margin. Trying to pin down why NCQ slows the Barracuda down in some tests is more difficult, although the Raptor with TCQ is also plagued by poor single-user performance, so overhead associated with the Promise FastTrak TX 4200 controller may be the culprit.
At the end of the day, Native Command Queuing is undoubtedly The Right Thing to do. It's clearly a winner for servers that face multi-user loads, but the implementations we tested need to mature before they become attractive alternatives for single-user desktops. Given that it's entrenched in the Serial ATA II spec, we'll be seeing a lot more of Native Command Queuing moving forward. I can only hope that the next generation of drives and SATA controllers will, across the board, at least be no slower with NCQ than they are without.
