[Beowulf] Barcelona vs. Woodcrest, computational chemistry research

Joe Landman landman at scalableinformatics.com
Thu Sep 27 19:35:40 PDT 2007



Guilherme Menegon Arantes wrote:

[...]

> This transformation will fit in core only for the very small problems,
> and usually you already need some decent disk-IO in this step. Actually
> depending on the size and details of calculation, matrices become so
> large that you need 100s of Gb, so fast disk-IO is a must.

A somewhat different calculation, coupled cluster, can also give 
hundreds of GB of IO per run.

Without meaning to sound like a commercial, we have provided high 
performance file servers with fast CPUs and lots of memory to handle 
some CFD jobs, specifically Fluent and alike, for some customers.  The 
performance difference is spectacular when you give them 10x the 
bandwidth to disk they had before.

Out of core is horribly slow (multiple orders of magnitude slower than 
in-core), but if you are going to do this, you may as well do this as 
quickly as you possibly can.


> The smaller systems/problems that would fit in core (and finish in a
> matter of seconds) would probably take minutes to finish doing a direct
> (or semi-direct) method (recomputing), so no one needs a fancy 
> workstation for them nowadays.

Yeah ... DFT !  I remember a waiting 1 week for 100 time steps, now my 
laptop can do the same calcs in minutes.

> 
> 
>> If you were working with larger systems and higher levels of theory you probably would have to resort to recomuting the integrals and therefore would be less concerned about bandwidth and more focused on clock and pure floating point.   Your probably job mix and benchmarking it will be key in making the right choice.
> 
> 
> Remember that scalling with most QC methods is very steep (from O(N**3) 
> to O(N**7))! So, depending on the size and level, even direct computations
> will not sufice and disk-IO will be very important as well. 

The CC codes go as O(N**10)-ish (not quite this; it is overly 
simplistic, it is more like O(N**6 M**4) or something like that).

Big, memory hungry, and eats IO bandwidth like it is going out of style.


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
        http://jackrabbit.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 866 888 3112
cell : +1 734 612 4615




More information about the Beowulf mailing list