The ABINIT Users Mailing List ( CLOSED )

[Nicola Marzari <marzari@mit.edu>]

Dear Xiulin,


again you might want to test. Our experience, per cpu or per core,
has always been that Intel was better - either because of
sheer GHz speed in doing blas/lapack operations, or
even with the current cores at lower clock because of the
excellent MKL library that contains all the blas/lapack subroutines.
We didn't specifically looked at Abinit, but I think plane-wave
electronic-structure codes are all similar. If anything, it matters
quite a bit how heavily the code is blas/lapacked, i.e. how much these
subroutines are used to do dense matrix algebra.

What AMD used to do much better was scalability from one to two
cores (or cpu) on one motherboard, due to the two buses to memory.

Not sure about the current generation of AMD machines, though.

For a cluster, I'm afraid you might need to test switches as
well (performance can be very different, at least for gigabit
ethernet) and MPI protocols.

E.g. for us LAM 7.0.6 is very inefficient on gigabit ethernet
and fast dual core Intel above 2 or 3 blades (2 or 3 depending
on the switch) while LAM 7.1.3 is much better, and we scale pretty 
linearly up to 6 blades/12 cores.

Some tests are in
http://quasiamore.mit.edu/pmwiki/index.php?n=Main.CP90Timings
comparing switches or mpi protocols. As you can see, the dgs1008tl
was significantly better than all other switches on lam 7.0.6
(well, you could scale well up to 3 blades, instead of 2), while with
lam 7.1.3 one scales well up to 6 blades (we didn't try more).

                                nicola



Xiulin Ruan wrote:
> Thanks for the information. We tested quad-cores and we do find the
> performance is no better than dual-cores. So we will go with dual-core
> processors. One more question. It seems AMD processors have better
> performance than Intel for abinit, right?
>
> Thanks,
>
> Xiulin
>  
>
> > -----Original Message-----
> > From: Nicola Marzari [mailto:marzari@mit.edu]
> > Sent: Saturday, August 18, 2007 2:50 PM
> > To: forum@abinit.org
> > Subject: Re: [abinit-forum] computer cluster for parallel computing
> >
> >
> >
> > Also, make sure you really need quad-cores. I've tested other plane-wave
> > electronic structure codes, and never found a case when the quad cores
> > were justified, for the marginal increase of performance compared to two
> > cores - you usually saturate the memory bus very quickly, even with one
> > core.
> >
> > This is even more relevant when you factor in the true costs of air
> > conditioning and electricity.
> >
> >                                 nicola
--
---------------------------------------------------------------------
Prof Nicola Marzari   Department of Materials Science and Engineering
13-5066   MIT   77 Massachusetts Avenue   Cambridge MA 02139-4307 USA
tel 617.4522758 fax 2586534 marzari@mit.edu http://quasiamore.mit.edu