[molpro-user] hardware suggetions for single Molpro workstation
Jonathan Dundas
jdundas at cm.utexas.edu
Mon Oct 24 17:41:50 BST 2011
Gerald and Kirk,
Thanks so much for your recommendations. We were able to spec out a workstation for less than $4K with a Sandy Bridge processor, 32GB of RAM and 4 600GB 10KRPM SAS drives.
Jonathan
----- Original Message -----
>
> I would also mention that for the scratch array where Molpro will
> write its temporary files, you certainly don't want raid5 since you
> want raw performance and don't care at all about redundancy. Simple
> raid0 is usually the best bet with at least 3 SATA drives. I would
> get a single (or multiple if you want to raid5 this) drive for the
> system and user files.
>
> -Kirk
>
>
> On Oct 20, 2011, at 10:42 AM, Gerald Knizia wrote:
>
> > Jonathan Dundas wrote:
> >> His description so far was that they plan on doing matrix
> >> manipulations and associated operations, and specified as many
> >> cpu cores & as much ram we could fit into a box for around $4k.
> >> With that price point & purpose in mind, and not knowing any
> >> better myself, a sample config we can afford will have:
> >>
> >> dual quad core Xeon E5620 2.4Ghz processors w/ 12MB L3 cache
> >> 24GB of RAM
> >> simple raid-5 of 3 500GB SATA 7.2KRPM drives
> >>
> >
> > Sounds good to me. If you can get some of the new Sandy Bridge
> > processors instead of older Xeons, which can do 8 double precision
> > flops per cycle instead of four, Molpro would also massively
> > profit from that (in many calculations this actually would give
> > you a performance boost of factor two).
> > Also, depending on the calculations, more disk space might or
> > might not be beneficial.
> >
> >> Now alternatively I saw that Molpro supports CUDA and that instead
> >> of trying to buy up a lot of RAM and processor instead I could
> >> skimp on that and get a Tesla C2050 card instead. What are the
> >> chances of it outperforming the above configuration with:
> >>
> >
> > Unlikely. CUDA support is still very preliminary and is not
> > supported for all types of calculations. In the current
> > development version the support has been extended, but in the
> > production version more CPUs and more RAM would be much better.
> > --
> > Gerald Knizia
> > _______________________________________________
> > Molpro-user mailing list
> > Molpro-user at molpro.net
> > http://www.molpro.net/mailman/listinfo/molpro-user
>
>
More information about the Molpro-user
mailing list