[snip]
The disk mechanism itself is not at issue as those are in principle
identical.
[snip]
This is certainly true, but omits the important fact that *none* of the highest
performance disk drives are available with an IDE interface. All the high end
disk drives also have sophisticated head motion optimization algorithms which
can substantially improve performance when the drive is given multiple commands
at a time to work on. The IDE interface has no way of even giving the disk
more than one command at a time. IDE looks good on paper for sequential I/O
performance, but not for random I/O on multitasking systems. Here's an example
of the measured performance of two SCSI disks for 512 byte block random I/O.
The Quantum Fireball Tempest is available in both IDE and SCSI; the SCSI
interface on this disk is poor, but doesn't affect the random I/O particularly.
The Quantum Atlas is SCSI-only and is even the previous generation (I don't
have an Atlas II). The Queue Depth is the number of concurrent commands given
to the drive.
Queue Depth = 1 Queue Depth = 15 Queue Depth = 31
================ ================ ================
Fireball Tempest 47 ops/sec 52 ops/sec 50 ops/sec
Atlas 71 ops/sec 129 ops/sec 151 ops/sec
As with all performance comparisons, what's crucial is the performance on the
workload that's important to the people making the hardware selection. If your
disk usage is not heavily multitasking and you don't use more than one disk per
channel concurrently, then IDE should equal 2nd-tier SCSI in raw performance,
if not in robustness (can IDE drives automatically reassign a bad block if an
error occurs? SCSI can). For any serious multitasking system with more disks
than IDE channels, SCSI is a big win even over bus-mastering EIDE.
Leonard