[rescue] Apple Announces Rackmount Server - WITHIDEHARDWARE RAID

[Patrick Giagnocavo 717-201-3366]

On Wed, May 15, 2002 at 09:19:58PM -0500, Phil Schilling wrote:
> Patrick Giagnocavo 717-201-3366 wrote:
> 
> > SCSI snobbery at its finest.
> > 
> > The guts of the drive are the same.  They come off the same production
> > line.  Some get IDE electronics and connectors, some get SCSI.
> > 
> But the controllers and subsystems are completely different.  Little CPU usage
> by SCSI and almost 100% by IDE,  multiple queueing of requests rather than
> single request at a time.

Ah, now I understand.

You have outdated information.  I know because I have run the tests
myself under controlled conditions using IOMeter and the Linux hdparm
utility.

Earlier modes of IDE did, in fact, suck.  But the later modes (circa
Intel TX mobo in I think 1995-96) addressed the issue.

Any time the transfer mode is UltraDMA2 (UDMA2) or higher, the method
of transferring the data from the drive is --exactly-- the same as the
way a SCSI controller transfers the data - DMA access directly to RAM,
then 1 interrupt to the CPU to let it know it is there.

The different on a P3-700 is less than 0.5% CPU utilization between
IDE and SCSI.

I understand that Solaris on the Sparc/IDE based platforms is not as
good, but this is not a question of the hardware but of the crappy
drivers Sun wrote for it - it would not surprise me if they
deliberately crippled it.  

For server throughput the answer will always be "multiple spindles"
and "seperate drive channels".  

Cordially

Patrick Giagnocavo
patrick at zill.net