[Wien] Intel(R) Xeon Phi™ coprocessor
Peter Blaha
pblaha at theochem.tuwien.ac.at
Thu Sep 12 08:56:19 CEST 2013
This depends a lot on what you want to do and how much money you have.
The single-core speed of a fast I7 is at least as good (or faster) than
most Xeons, and they are MUCH cheaper. So for all systems up to 64-100
atoms/cell, where you need several k-points, a small cluster of I7 cpus
/4 cores, or more expensive 6 core) (with GB-network and a common NFS
file system) is for sure the fastest platform and in particular has by
FAR the best price/performance ratio (One powerful I7 computer may cost
about 1000 Euro). For bigger clusters, a drawback can be the large
"space" to put all PCs on a big shelf ....), but if you have less than
10000 Euros, this is probably the best choice.
However, Xeons can be coupled (2-4 Xeons) to a "single multicore
computer (eg. 16 cores)", which may work with mpi and can be used to
handle systems up to 200-300 atoms. They also can be bought in small
boxes and may fit in a single 19 inch cabinet. But of course such
systems are much more expensive. From what I said above it should be
clear, that it is completely useless to buy a "single 4 core Xeon computer".
The next step would be to buy an Infiniband switch+cards and couple your
PCs with this fast network to a powerful multinode mpi-cluster. Since
the switch/cards are fairly expensive, on usually takes here Xeons as
platform. However, you need to know how to install/configure the
software properly. I've seen such clusters even in computing centers,
which were completely useless, because the network/mpi was instable and
jobs would crash randomly every couple of hours .....
Our strategy:
i) We have a GB-networked cluster with Intel I7 computers (which we
maintain our-self and this cluster includes also all the user-
workstations) and do all the calculations for systems up to 64
atoms/cell on these systems.
2) For bigger systems we go to our University computer-center and run
there with a PBS queuing system. This has the advantage that we do not
need to care about the installation of the infiniband network nor the
mpi-infrastructure (but we use always intel-mpi together with ifort/mkl).
On 09/11/2013 06:31 PM, Luis Ogando wrote:
> Dear Prof. Blaha,
>
> Just for curiosity, what processor did you buy ?
> Is the Xeon family better than the i7 one for WIEN2k calculations ?
> All the best,
> Luis
>
>
> 2013/9/11 Peter Blaha <pblaha at theochem.tuwien.ac.at
> <mailto:pblaha at theochem.tuwien.ac.at>>
>
> I don't know what "latest" means. We use the latest one installed on
> our supercomputers (4.1.1.036)
>
> I have not seen any significant change with mpi in the last years.
>
> PS: I just got info that we have now a new ifort available for
> download ...
>
>
> On 09/11/2013 05:00 PM, Laurence Marks wrote:
>
> Thanks.
>
> One thing I will add/ask concerning the parallelization, the latest
> impi seems to be substantially better -- have you tried it? I
> have not
> just noticed this with Wien2k, but I am told that others have seen
> improvements in other codes.
>
> On Wed, Sep 11, 2013 at 9:42 AM, Peter Blaha
> <pblaha at theochem.tuwien.ac.at
> <mailto:pblaha at theochem.tuwien.ac.at>> wrote:
>
> Before buying a couple of new computers, I was asking myself
> the same
> question and discussed this with some people of our
> computing departments.
>
> The conclusions:
> a) potentially very good, but in practice very questionable,
> because for
> most application you cannot get out the real speed (10 times
> faster than
> an Intel I7). This is true, even for many lapack/mkl
> subroutines where
> it "should" work better.
> They told me to "wait", until the mkl becomes better
> (hopefully). I'm
> not too optimistic, when you see how badly the
> mkl-parallelization of
> multicore machines is working (2 cores is very good, but 4
> or more is
> already very bad).
>
> b) The nature of our problem (big eigenvalue problem): A "fast
> processor" is useful only for large problems --> large memory.
> You can buy Phi coprocessors now with quite some large
> memory, but then
> they are terrible expensive (and 5 "normal" PCs are faster
> and cheaper)
>
> c) the hardware design has a VERY slow communication between
> main-memory
> and Phi-memory. This makes also parallelization over several
> PHI-nodes
> via mpi not really possible (if you need any significant
> data transfer,
> like for an eigenvalue problem).
>
> Thus I did not buy it.
>
> However, if anybody has access and time to try out WIEN2k on
> PHis, I'd
> would be very interested in getting feedback. (Maybe these
> computer-people were not good enough ....)
>
> PS: I know from G.Kresse that they had some time ago (maybe
> 2 years ?)
> an expert from NVIDIA with them. After 2 weeks of porting
> VASP to these
> GPUs by this expert, VASP on the GPU was "almost as fast" as
> on an Intel
> I7 processor.
>
>
> On 09/11/2013 04:16 PM, Laurence Marks wrote:
>
> Anyone know if these will be viable with Wien2k (mpi,
> i.e. large problems)?
>
>
> --
>
> P.Blaha
> ------------------------------__------------------------------__--------------
> Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060
> Vienna
> Phone: +43-1-58801-165300 <tel:%2B43-1-58801-165300>
> FAX: +43-1-58801-165982 <tel:%2B43-1-58801-165982>
> Email: blaha at theochem.tuwien.ac.at
> <mailto:blaha at theochem.tuwien.ac.at> WWW:
> http://info.tuwien.ac.at/__theochem/
> <http://info.tuwien.ac.at/theochem/>
> ------------------------------__------------------------------__--------------
> _________________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.__at
> <mailto:Wien at zeus.theochem.tuwien.ac.at>
> http://zeus.theochem.tuwien.__ac.at/mailman/listinfo/wien
> <http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien>
> SEARCH the MAILING-LIST at:
> http://www.mail-archive.com/__wien@zeus.theochem.tuwien.ac.__at/index.html
> <http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html>
>
>
>
>
>
> --
>
> P.Blaha
> ------------------------------__------------------------------__--------------
> Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
> Phone: +43-1-58801-165300 <tel:%2B43-1-58801-165300>
> FAX: +43-1-58801-165982 <tel:%2B43-1-58801-165982>
> Email: blaha at theochem.tuwien.ac.at
> <mailto:blaha at theochem.tuwien.ac.at> WWW:
> http://info.tuwien.ac.at/__theochem/
> <http://info.tuwien.ac.at/theochem/>
> ------------------------------__------------------------------__--------------
> _________________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.__at
> <mailto:Wien at zeus.theochem.tuwien.ac.at>
> http://zeus.theochem.tuwien.__ac.at/mailman/listinfo/wien
> <http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien>
> SEARCH the MAILING-LIST at:
> http://www.mail-archive.com/__wien@zeus.theochem.tuwien.ac.__at/index.html
> <http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html>
>
>
>
>
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
> SEARCH the MAILING-LIST at: http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html
>
--
P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-165300 FAX: +43-1-58801-165982
Email: blaha at theochem.tuwien.ac.at WWW:
http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------
More information about the Wien
mailing list