[Wien] system comfiguration

[L. D. Marks]

We went through a similar analysis a couple of months ago, so a few
comments about what we decided and (perhaps more important) why.

Questions to consider up front:
1) Are you going to buy complete systems or build from components? I was
fortunate to have students who could build, which gives more flexibility/$.
2) How much do the nodes cost. We considered this both in terms of current
costs, and what a CPU update in 1-2 years would be.
3) Which compiler is less flakey, and which one will be best 1-3 years
from now and where will good blas/lapack come from?
4) What operating system are you going to use?
5) How much does mpi matter?
6) Do you have to "fight" purchasing agents, and are there any overhead
issues?

When we went through the analysis, in terms of building systems if you
have to buy the box, power supply etc it turned out that going for the top
CPU's in most cases was cost effective.

One question is single-CPU versus dual-CPU. AMD seem to have the edge
with dual-CPU systems. However, with AMD you have to either struggle
to use ifc or purchase a compiler. At least as important, Wien2K seems
to spend most of its time in blas/lapack routines and the mkl libraries
seem to be rather good (and continue to improve). Although there are the
goto libraries, how long will they continue to be developed? We decided
that (despite its problems) ifc/mkl is probably here to stay and will only
get better, which rules out AMD. Since lapw1 seems to be the slowest step,
and this is parallelized by shipping calculations off to different
computers, mpi does not seem to be that critical (perhaps others can
comment). Note: I don't know if other code such as VASP will benefit from
dual-CPU machines.

We therefore decided to go with 3.4GHz P4 with dual-memory. (Supposedly
these are hardware compatible with future CPU's so simple CPU replacement
1-2 years down the line would be possible.) We did not find a
cost-effective way of going beyond 2G memory, so settled for this.
We put in a 1G ethernet to connect them, which may not be that critical
(but in the future might be useful). The benchmark (using -mp1 -prec_div
rather than -mp) is ~ 212 seconds with either ifc7.1 or ifort8.0 and
mkl6.1 (mkl6.1 is faster than mkl6.0) and is on the Wien web page.

There are some teething problems. Ifort8.0 does not seem to want to
compile everything right, probably over-optimizing (needs a little work).
We also thought we got Prescott's, but cannot compile with this option or
any of the P4 options with ifc7.1, no idea why. Some of this may be
related to having Red-Hat 9.0 which has compatibility issues with ifc.

I hope this is useful -- it's not a straight answer. Remember, what ever
you buy will be obsolete by the time it arrives.

-----------------------------------------------
Laurence Marks
Department of Materials Science and Engineering
MSE Rm 2036 Cook Hall
2225 N Campus Drive
Northwestern University
Evanston, IL 60201, USA
Tel: (847) 491-3996 Fax: (847) 491-7820
mailto:L-marks at northwestern.edu
http://www.numis.northwestern.edu
-----------------------------------------------