On Fri, Feb 23, 2001 at 09:18:34PM +0100, Manuel Bouyer wrote:
> > 
> > What is it that causes the difference in CPU load between IDE and SCSI?
> > IDE uses DMA just like SCSI, so...  Tagged queuing?
> 
> With IDE drives you've at last one interrupt per xfer. With a decent
> SCSI adapter and driver you should be able to have serveral xfers per
> interrupt. Also with today's IDE controllers you still have to read/write
> registers on the PCI bus, before and after each xfer. This is slower than
> main RAM, of course. With good SCSI controllers read/write of registers
> per xfer should be much less (possibly less than 1 per xfer, as for
> interrupts).

There's another really significant difference that means that using a
large number of IDE drives on traditional "pciide" controllers in the
same machine is a poor idea:  because the controller is totally
unintelligent and has no local buffering, *any* contention on the PCI bus
will dramatically increase latency and reduce bandwidth.  With as few as
two 7200RPM drives required to saturate the PCI bus given typical maximum
latency values, and since there are no pciide controllers with more than
a 32bit, 33-MHz bus interface, it's trivial to see this effect in action.

For most workloads, a six-drive IDE lash-up will dramatically underperform
a three-drive or four-drive one.  Yay.