Hi Ewan,
I think this can only be true if the cost of the RAM etc. as
fraction of the system cost is greater than the % gain due to HT , whereas
the real world numbers are ~10% and ~50% ?
Also, for I/O jobs, once you go to 10GbE, you generally want to max out
the performance of the individual boxes to try to match the available bandwidth
because the per port costs are still expensive.
Alex
On Friday 09 Mar 2012 18:07:31 you wrote:
> _______________________________________
> From: Testbed Support for GridPP member institutes
> [[log in to unmask]] on behalf of Rob Fay [[log in to unmask]]
> Sent: 09 March 2012 15:46
> To: [log in to unmask]
>
> > Consequently, we intend to initially run with 16 jobs per node and see
> > how that performs in real life.
>
> How much memory and disk have you got? A job slot isn't just about the CPU.
> What's always put me off hyperthreading is that while there's clearly a
> performance benefit, the cost of the extra RAM etc. that you need (twice
> as much if you use all the HT units) makes that extra performance very
> expensive as compared with just buying more real CPUs.
>
> The only time we've ever used it was when we borrowed some local cluster
> nodes that had 4Gb of memory per core, so still having 2Gb per HT.
>
> Ewan
--
------------------------------------------------------------------------------
| |
| Dr. Alex Martin |
| e-Mail: [log in to unmask] Queen Mary, University of London, |
| Phone : +44-(0)20-7882-5033 Mile End Road, |
| Fax : +44-(0)20-8981-9465 London, UK E1 4NS |
| |
------------------------------------------------------------------------------
|