1/ Most web sites I saw, report single-user performance as a pattern of hard disk drive IO. They try to reproduce common tasks of a single user, whether there are one or many drive in the benchmarked system, and measure the elapsed time spend in IO retrieving.
This pattern typically consists of hard drive access with good data proximity and a short queue of user requests to the IO subsystem; it makes drive sustained transfer rate shine. It is often opposed to a multi-user pattern where data proximity is variable and queues of user requests are deep. This multi-user pattern is often associated to a "server pattern".
2/ NCQ is not only about drive performance: it allows, if supported by the OS, the hard drive interface controller, and the drive itself, to deliver IO commands some quality of service (a tag) and fast, drive-controlled and initiated data to memory transfer.
To my knowledge today, Windows doesn't offer a NCQ enabled driver stack (it will in Vista, ATA-port, but I don't have any info about the envisaged implementation), and Linux offer some test driver to play with. Currently under Windows it could be implemented as a SCSI-port or STOR-port driver for the hard drive controller in which general CQ are wrapped around NCQ functionalities.
In the hard disk controller interface (the chipset or discrete SATA controller chip), prior to the AHCI specification, NCQ implementation was ambiguous and often not interoperable between drives/controllers.
In the drive, NCQ is implemented as support for the command set. Part of this implementation could employ a reordering command algorithm to maximize the drive throughput at the expense of latency of some commands. But NCQ in itself doesnâ€™t necessitate the use of this algorithm.
What the current crop of benchmark tries to measure is this part of NCQ: the ability for a drive to reorder effectively commands to maximize throughput.
And in effect, for a particular command, this algorithm, when implemented, could be slower than using no reordering algorithm: imagine retrieving two pieces of data on the disk, the reordering algorithm objective is to make sure retrieving these two is faster but the time to retrieve the first of the two could be longer. Thatâ€™s why NCQ enabled drive today shine more on multi-user benchmark.
The full benefits of NCQ regarding performance and other aspects would only be unleashed by a good software implementation, a good controller implementation and a good drive implementation and as stated, it isnâ€™t the case actually.
3/ WD implement in some Raptor (some models firmware have it others not) a command queuing algorithm said to mimics TCQ (Tagged Command Queuing, an SCSI command queuing scheme) but there's not much information about what part of it is in the box and it will suffer similarly from the problem explained above.