[Gllug] [Fwd: Server]

[Mike Brodbelt]

On Sun, 2003-11-23 at 14:38, Richard Jones wrote:
> On Sun, Nov 23, 2003 at 02:15:35PM +0000, Mike Brodbelt wrote:
> > Also bear in mind that SCSI electronics are much better than IDE for
> > servers with lots of disks (no siginificant per disk CPU loading like
> > you get with IDE).
> 
> I'm assuming you're talking about ancient IDE disks running in PIO
> mode?

Even modern disks in UDMA5 mode have a higher load on the host CPU than
SCSI disks.

>  These days both do DMA, both are command driven, and, heck, both
> use similar logical block addressing structures.

There's a large debate about "which is better" here that I don't intend
to get into in any depth, as it can go on for ages. That said, I do feel
the need to stick my oar in to a small degree :-). There's an
interesting article about one person's experience here:-

http://hardware.devchannel.org/hardwarechannel/03/10/20/1953249.shtml

This isn't really a scientific test, but the followup threads are worth
reading, and show both sides of the argument.

> I'm betting that manufacturers are probably using exactly the same
> electronics, and just changing the firmware and connector/IO side.

I think the differences are significantly greater than that. A few real
world considerations that come to mind:-

* Tagged Command Queuing is still not that common in IDE drives, and I
think I remember reading the linux implementation had a few problems
with TCQ on IDE drives. 

* The IDE bus isn't even close to U320 performance, which is an issue
with multi drive systems. 

* The 2 drive/channel limit on IDE is significant for servers. 

* Real world SCSI controllers for servers are far higher spec than
available IDE ones. 

* Sustained data transfer is excellent on IDE due to high areal
densities, but SCSI disks usually outperform on the workloads that
typical servers perform.

* The highest performance drives are only available with a SCSI
interface. To the best of my knowledge, you still can't get 15000rpm and
8Mb cache on and IDE disk.

* Operating characteristics of SCSI drives tend to show higher MTBF and
higher operational temperatures. Looking at Seagate's website, the
Cheetah SCSI disks are listed with a 1,200,000 hr MTBF - the ultra ATA
100 drive data sheet doesn't even list MTBF.

Comparing a single IDE drive to a single SCSI drive does tend to flag
SCSI as more expensive, for little advantage. When you're building
redundant servers with >5 drives though, and you're looking for a
storage system, not just a drive, the story is different. That's why
when you buy a 1000ukp PC it comes with an IDE drive, and when you buy a
10,000ukp high end server, it comes with SCSI. If the OEMs genuinely
believed there was no difference, your 10,000ukp server would have IDE,
and they'd have bigger profit margins...

Mike.


-- 
Gllug mailing list  -  Gllug at linux.co.uk
http://list.ftech.net/mailman/listinfo/gllug