Actually, a lot of it depends on the architecture...
Xeons are really only designed for 4 way SMP.. 8 and 16 way solutions are
custom designs and the cost per CPU goes way up.
Plus, the GTL bus that Xeons use can easily saturate the memory bus at the 4
way level.. All 4 CPUs have to share the same memory bandwidth. So, they
dont scale nearly as well as Sun, Alpha or MIPS.
The Alpha uses the EV6 bus.. It is quite capable of scaling well to 16 and
32 CPUs with little problem.. But, the Alpha is a high-cost option as well.
Sun UltraSparc II and III processors can scale as high as 64 processors in a
single box.. SGI has Origin servers that can scale even higher..
Linux will run well on any of these platforms..
Also, for a better high end disk solution, you might try checking into EMC
hardware.. They are expensive, but they're about the best around for Unix
servers.. IBM is the best for mainframes... If you're just looking for a
small solution (say under 200 gig) you may want to stay with a
Mylex/Compaq/Adaptec RAID soluiton. But, anything beyond that is where EMC
sits..
My employer has used EMC on our cluster of Sun E10000 systems, and has been
quite pleased.. We have 4 or 5 terabytes for our Unix machines. They run
24/7 with no problems.
----- Original Message -----
From: "Rusty Carruth" <[EMAIL PROTECTED]>
To: <[EMAIL PROTECTED]>
Cc: <[EMAIL PROTECTED]>
Sent: Wednesday, August 08, 2001 5:14 PM
Subject: Re: [expert] SMP systems (continued)
> "Julia A. Case" <[EMAIL PROTECTED]> wrote:
> > Thanks to all that sent me email about multi cpu mother boards... One
> > more question... I've used linux quite a bit with dual CPU systems, but
> > I'm thinking about a motherboard that supports 4 CPU's... I've heard
> > rumor that you just don't get a linear increase in computing power when
> > you go over 2 CPU's...
>
> True, but you don't get linear from 1 to 2 either.
>
> And the rumor I'd heard was that you don't lose much of the multiplier
till
> you reach like 16 cpus.... (If thats really just a rumor I'm sure we'll
> get corrected here real soon now ;-)
>
> > That the task scheduler doesn't make efficent use
> > of the extra CPU's...
>
> That's a new one on me. Maybe thats true of the 'other' OS, but that's
not
> my understanding on linux.
>
> > Also is it possible to do something like dedicate
> > a single CPU to just doing filesystem I/O?
>
> No. Then it would not be *S*MP...
>
> > That would make using
> > software RAID almost as fast as using a hardware RAID controller.
>
> Actually, since Linux can (I think) share KERNEL tasks across multiple
cpus
> as well as the user tasks, that you get this bonus by using 2 cpus
anyhow....
>
> Again, standing by for correction ;-)
>
> rc
>
>
> Rusty E. Carruth Email: [EMAIL PROTECTED] or
[EMAIL PROTECTED]
> Voice: (480) 345-3621 SnailMail: Schlumberger ATE ___
> FAX: (480) 345-8793 7855 S. River Parkway, Suite 116 \e/
> Ham: N7IKQ @ 146.82+,pl 162.2 Tempe, AZ 85284-1825 V
> ICBM: 33 20' 44"N 111 53' 47"W http://tuxedo.org/~esr/ecsl/index.html
>
>
