[Gllug] Best option for a lot of compute power

[John Hearns]

On Mon, 2004-06-07 at 10:49, Daniel P. Berrange wrote:

> > We sell a lot of opteron kit, you can get a dual opteron box for not
> > all that much money which'll kick the pants off a dual p4 doing
> > hyperhtreading.
> 
> One of the real killer features of the Opteron is its memory
> bus architecture which lets the memory bandwidth scale up
> linearly as CPUs are added. By constrast no matter how
> many Xeon CPUs you've got they all share the same fixed 
> memory bandwidth. So for applications which are memory
> bandwidth limited, going from 2 Xeon to 4 Xeons adds nothing,
> but going to 4 Opterons absolutely kicks ass.

I agree.
Extend that with Infiniband as the interconnect between nodes, mmm......

As regards hyperthreading, we normally switch it off on compute nodes.
For HPC type applications, the concensus seems to be it has no
advantage. But as always YMMV - worth benchmarking with/without HT with
your applications.

-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug