[excuse the cross-post to LHCb and LCG]
Hello everyone,
I have a "data point" I'd like to share with you, regarding 64-bit
architectures.
[log in to unmask] wrote:
> On Thu, 11 Nov 2004, Andreas Unterkircher wrote:
>>We're thinking of compiling on IA64 not running the IA32 binaries on IA64
>>(this would be possible in a slow emulation mode but i think it wouldn't
>>make much sense).
>
> Given that even Intel now makes CPUs which implement AMD64, and HP
> has stopped selling Itanium (at least, that is the rumour I heard
> -- and remember, HP are the ones that did most of the R&D behind
> Itanium) is this really something we will benefit from in the
> future?
I have spoken to about 10 different people on the subject of 64-bit
architectures this week at Supercomputing, either from large US
computing centres, or from cluster manufacturers (Penguin, Rackable, and
several others). They all report essentially the same thing:
1. Hyper-threading should be turned off everywhere.
2. AMD 64-bit architecture is up to 30% faster than Intel, thanks to
much faster direct memory access, except
3. Intel Itaniums currently have PCI-X support, which gives them some
improved device access, however
4. The next generation of AMD 64-bit processors will support PCI-X
The cluster manufacturers still sell 70%+ Intel processors, but most of
them admitted they thought this was because "No one gets fired for
buying Intel". Within large research computing centres, as one vendor
put it, "No one even talks about Intel" (a bit of an overstatement, in
my opinion).
Anyway, the *very* clear message seemed to me that AMD 64-bit opterons
are going to be the future of scientific computing clusters, and that if
you have any Intels, turn off Hyper Threading (all the cluster vendors I
asked about it said "We turn it off by default").
I really wanted even just one person to tell me the benefits of Hyper
Threading (Intel included) in a computing cluster environment, and they
couldn't: if you have a web server, commercial transaction server, it
might help you out, but for scientific computing you are much better off
without it, and, if necessary, just running two processes (jobs) on the
same real processor is a better bet.
Regards,
Ian
--
Ian Stokes-Rees [log in to unmask]
Particle Physics, Oxford http://www-pnp.physics.ox.ac.uk/~stokes
|