See also: Promise FastTrak ATA RAID controller.
A while back, the Storage Review took a look at Promise's FastTrak, an ATA RAID controller that provides a viable alternative to more expensive SCSI adapters in striping and mirroring applications. In those tests, we found that the controller in RAID 0 (striping) configuration provided a dramatic increase in sequential data transfer (STR). Though this translated into a relatively modest performance increase in WinBench 98's high-level application tests, such an increase in STR is a boon to those performing non-linear A/V editing.
Windows NT Workstation 4.0 features a built-in function that allows spans of spaces on two different drives to be used in a striped configuration. Though NT's Disk Administrator is actually pretty flexible in allowing the user to create a striped pair in a variety of setups, the ideal configuration is the complete striping of all space between two identical drives.
One does run into some logistical problems when running an NT software-striped pair when using ATA drives. First off, a system can't be booted off of a striped partition. This makes sense, of course, since OS support of a function needs to be "primed" by booting up the OS itself. Thus, if two identical drives are going to be wholly devoted to the striped pair, NT must be installed on a third boot drive. As we all know, motherboards today have only two ATA channels. In our testing of drives, we always have the boot drive connected to the primary controller while the test drive resides by itself on the secondary channel. Just simple insurance to minimize the effect of contention between two drives on the same channel. So, how does one set up a software-striped drive pair? At all costs, avoid setting up the two drives within the stripe on the same channel- you'll lose virtually all the promised performance gain. We ended up with one of the pair as a slave on the primary controller and with the other drive as the master on the secondary channel. Though it wasn't good for my piece of mind, such a setup should not limit performance within the context of benchmark that's running tests -only- on the striped pair. When running a setup in real-word usage, however, the user should take care to minimize access of the boot drive- all applications should be installed to the striped pair as should NT's swapfile. Of course, all this headache could be avoided in several ways. An all SCSI setup (SCSI boot drive and two identical SCSI units in the striped pair) wouldn't be subject to the same limitation. You could add more ATA channels also, though if you're going to go out of your way to buy a controller and give up system resources for it, you may as well just use a FastTrak.
For the RAID, we decided to use a pair of DiamondMax Plus 2500's, the top ATA performer in Windows NT. Formatting the stripe did result in a problem. Whenever we test drives under NT, we partition the test unit as an extended volume. We've never run into problems formatting even the largest drives with such methodogy. Setting up a primary partition on a test drive larger than 8.4 gigs, however, usually results in a failed format. Since we did not get the option to choose between "primary" and "extended" when setting up a striped pair, attempts to format the drive always failed. It turns out that there's a patch that can fix this problem at http://support.microsoft.com/support/kb/articles/q183/6/54.asp. Thanks to Dagfinn R. for the pointer. Using this patch, we were successfully able to format the single 20GB partition both with NT's software striping and with the Promise FastTrak (results presented for comparison).
I have to admit that I went into these tests with a preconceived notion that hardware-based striping would outperform NT's software function, in CPU utilization if not anything else. Surprisingly, however, software-striping mananged to keep pace with the FastTrak in all categories. Further, software striping ended up actually being slightly -less- CPU intensive per kilobyte transferred. In all high-level tests, the software came within 2% of the figures posted by the FastTrak controller. In terms of low-level sequential data transfer, software-striping topped out at 26.6 MB/sec. The Promise adapter, on the other hand, turned out a true doubling of a single drive's score, 29.8 MB/sec.
So, in terms of raw performance, NT software-raid striping is just as fast as hardware-based RAID 0. Again, however, it is less flexible, especially when working in the confines of the ATA interface. With a pair of SCSI drives, it's a boon-- SCSI's natural flexibility means one gives up relatively little to run the drives in an NT striped pair. In an ATA system, however, implementing software striping means that you're going to give up the "pristine" setup of having only one device per channel. If on top of all of this you're also using an ATAPI CD-ROM, it all gets quite nasty in my opinion. In such cases, the flexibility of adding the Promise FastTrak still makes it quite a compelling option. The two striped drives can then reside on the single-IRQ FastTrak (which also allows you to boot off of the striped drives) and the ATAPI CD-ROM can remain on its motherboard-based channel. The boot drive used with software striping would no longer be necessary. All in all, the FastTrak provides a cleaner, faster setup than the jumble you'd end up with when using 3 ATA drives and an ATAPI CD-ROM on two ATA channels.