[Beowulf] cpu/mobo recommendations

[Andrew Latham]

I belive that the Archives for this year contain tons of advice to this tune.

To recap what I have read.

-Mobo and chips - bleeding edge can bleed you, Heat is the issue,
Quads are an option, SMP _can_ require uncommon ram setup on new
boards.

-Pick you case early.  - few screws, good vents, tested PS

-Request test units from vendors - just to see them

-Talk to Manufacturers. (example Tyan guys read this list)

-Diskless units have scale issues - VLAN them with bridges + multi
servers read as
"Do your network homework twice, once for startup and again for operation"

-Don't over look buying the hardware preconfigured

-Good server room means a good home - have plenty of AC and access control

-Read

-Be happy with what you choose.

Robert will fill in any gap I left.........


On Fri, 24 Sep 2004 09:36:57 -0500 (CDT), Bill Comisky
<bcomisky at pobox.com> wrote:
> 
> It's time to upgrade the cluster, and I'm trying to determine the
> best hardware configuration for our purposes.  It's been a while since I
> went through this exercise and there seem to be more choices now (the last
> time I added nodes, people were just starting to use AMD clusters).
> 
> Specifically, the nodes will be diskless, and running coarsely grained
> parallel problems distributed by Sun's Grid Engine.  So fast ethernet
> works fine for networking.  Also the distributed jobs are Fortran77/90 and
> C codes, which I can recompile for 64bit processors if need be.  My main
> criteria are:
> 
> - best price/performance ratio:  our simulations are essentially
>    independent, so overall runtime for a set of jobs will scale almost
>    linearly with number of processors.  So the best solution is likely not
>    the cutting edge fastest, but a step back from the fastest where we can
>    buy more nodes for the same $$.
> 
> - hardware reliability
> 
> - rack form factor
> 
> I'm assuming 2 way SMP motherboards are still the way to go.  If anyone
> has gone through the same process recently, I'd appreciate any
> recommendations or pointers to good info.
> 
> thanks,
> bill
> 
> --
> Bill Comisky
> bcomisky at pobox.com
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> 



-- 
Andrew Latham

http://www.lathama.com
lathama at gmail.com
lathama at yahoo.com
lathama at lathama.com
If any of the above are not working we have 
bigger problems than my email.