[rescue] Wow. I feel dumber..

Patrick Giagnocavo +1.717.201.3366 patrick at zill.net
Sun Jan 19 16:45:08 CST 2003

On Sun, Jan 19, 2003 at 03:13:21PM -0500, Dave McGuire wrote:
> On Sunday, January 19, 2003, at 12:12 PM, Patrick Giagnocavo 
> +1.717.201.3366 wrote:
> > I suspect that the IDE drivers for Solaris are just generally sucky.
> >
> > There should be little difference in performance between a 7200rpm IDE
> > drive and a 7200rpm SCSI drive, but on the IDE based Sparcs, there is;
> > thus the driver in general is not that great.
>    I suspect a lot of that comes from the amount of processing that IDE 
> doesn't offload from the host.

I won't say you are wrong, because I don't know all the details of how
they do it on the Blade series.  

But IDE, per se, does not signicantly increase CPU load vs. SCSI, as
long as you are using "UDMA mode 2" or higher.  I know this because I
have tested it (when I was consulting at a large HDD maker): the
numbers between SCSI and IDE were within 0.5% CPU usage of each other
(5% vs. 4.5% on a PIII-800).

Previous IDE problems were that it did not use DMA to put the results
of a disk query into RAM, and instead generated 1 interrrupt per 512
bytes transferred, which the CPU would have to handle; so it would
stop what it was doing, transfer 512 bytes, then switch back to its
other tasks.  This resulted in a large number of interrupts which, on
x86, resulted in high CPU usage.

So, if the Sun-supplied driver uses the "old" way of doing it, they
are killing performance with interrupts.  Or maybe they do the DMA
transfer in an inefficient way, or the part of the chipset that
handles DMA is sucky.

The main advantages that SCSI vs. IDE offers is scatter/gather support
and that IDE can't usefully support more than 1 relatively busy device
per channel.  (Then again, you can easily buy a 2 or 4 channel IDE PCI
card for like $30.)

Say, Dave, did you ever get those T-cards or whatever they are called
produced so we can all use cheap FC drives?



More information about the rescue mailing list