[SunRescue] Ahh the old SCSI vs IDE thing, was Re: HBA's

[James Lockwood]

On Mon, 4 Dec 2000, Chris Byrne wrote:

> I've seen a lot of U30's on EBay, some of them for as low as $750, but I
> wont touch a one with less than 256megs of RAM, and at least one 4 gig
> disk, in perfect condition, and from a seller with a good reputation. 

I prefer barebones systems myself.  The U30's in question were also
waranteed by a commercial seller for 60 days, a period which I find
sufficient to work out any kinks which may develop.

Perfect condition is nice to have but not always necessary.  I had for a
while an SGI Crimson/RE which was backed into by a car (ok, a tank.  '68
Dodge Coronet wagon).  Looked like hell, frame was mangled and the machine
was totally unsellable at that point, but it never stopped working for a
minute. 

(The eventual fate of this machine was somewhat more severe.  I traded it
to a friend who wanted to place it in an environmentally controlled
computer room, so he showed up with a pickup truck.  He got it to the
facility, pulled it over a short plywood ramp to the loading dock, and
went off to find someone with keys to the back entrance.  A maintenance
worker bumped into it with a cleaning cart, started it rolling across the
dock on those nice smooth casters, where it went off the edge, fell
straight down 4 feet, and exploded into a huge pile of vaguely
crimson-colored parts.  I'm willing to bet that if it hadn't been for the
earlier damage to the frame, the machine would have survived the drop.)

> When comparing single drive to single drive on the same system, a high
> quality UDMA device will often best it's SCSI counterpart in
> straightline benchmarks. These are indicitave of very good single device
> single task performance. Change the configruation to two drives in each
> system and the picture changes significantly in favor of SCSI. Four
> drives skews the equation even more, because those four SCSI drives can
> deliver their full sustained bandwidth all at the same time (each
> typically higher than an equivalent IDE device), whereas the IDE devices
> will each burst out their first 2 mb's, then each one will deliver maybe
> 10mb/s sustained througphut, and it doesnt add together because bus
> mastering for IDE devices only works for one device at a time. 

This is very true, but it all pales compared to the differences between
the IDE driver as implemented in Solaris/SPARC (dad driver) and in
Solaris/x86 (ata driver).  The two have very different performance
characteristics under load, and I find the dad driver in particular to be
lacking when it comes to handling many small requests.

-James