Dear Alex, dear Drew,

I agree with Drew's statements below, and only wish to make a small
correction/addition. As soon as we have a stable ADF version with the Intel
compiler (the Intel people claim to have very recently fixed a compiler bug
that prevented us from proceeding with our tests), we expect that the
speed-up for PIII and Athlon systems will be about 25% w.r.t. to the
current ADF executables, but about 75% (but depending strongly on the
type of run and type of molecule, basis set etc.) for PentiumIV systems.
If all proceeds as we hope, this would be a strong point in favor of
PentiumIV at this moment.

ADF should be able to make use of the SSE2 instructions on PIV. It seems
that it indeed does, but this will have to be verified as soon as we can
make a stable version that passes all our reliability checks.

Best regards,
Stan

> Hi Alex,
>
> > I am planning to build or buy a 4-cpu linux cluster mainly to run ADF
> > and
> > Gaussian 03. I am wondering if anybody on this list can share some
> > experiences regarding the use of Athlon or P4 based PCs for that
> > purpose,
> > or whether there is some sort of benchmarking available for the two
> > processors. From what I gather on the internet, Athlon XP still seems
> > to
> > have the better FPU power, unless one can utilize the P4's SSE2
> > extension
> > to its instruction set. IIRC SSE2 is supported in the more recent linux
> > kernels, but is it utilized in ADF and does ADF actually run faster on
> > a
> > P4 compared to an XP? Due to budget limitations I am currently looking
> > at
> > either a Xeon P4 system or four individual P4 or Athlon PCs strung
> > together with 1 GB/s ethernet.
> We have an Athlon based system, and it caused some real headaches in
> terms of getting it stable. My feeling is that the quality of the
> Pentiums is better in this respect, and I think it is worth taking into
> account.
>
> Regarding performance, ADF is currently being tested with the Intel
> compiler, and results are quite promising. Performance seems to be on
> the order of 30% faster, if I recall properly. This compiler is
> optimized for Pentium chips, of course, and there is a reasonable
> chance it will be supported for ADF in future. I think this will give
> the Pentium the edge performance wise over the AMD.
>
> >
> > What is actually the bottleneck in a typical ADF calculation, say a
> > geometry optimization of a 50-100 atom molecule? Is it processor speed,
> > memory bus speed, disk IO, ...?
> Typically you get close to 100% CPU load, so I would say the bottleneck
> for ADF is the processor. It isn't disk IO or interprocessor
> communication for the most part. An exception to this rule is Band:
> Band is disk bound in a big way.
>
> Regards,
> Drew
>
> ========================================
> Dr. Drew McCormack (Kmr. R153)
> Afd. Theoretische Chemie
> Faculteit Exacte Wetenschappen
> Vrije Universiteit Amsterdam
> De Boelelaan 1083
> 1081 HV Amsterdam
> The Netherlands
>
> Email cormack_at_chem.vu.nl
> Telephone +31 20 44 47623
> Mobile +31 6 483 21307
> Fax +31 20 44 47629

------------------------------------------------------------------------
Dr. S.J.A. van Gisbergen SCIENTIFIC COMPUTING & MODELLING NV
Phone: +31-20-4447626 Vrije Universiteit, Theoretical Chemistry
Fax: +31-20-4447629 De Boelelaan 1083
E-mail: vangisbergen_at_scm.com 1081 HV Amsterdam, The Netherlands
                               http://www.scm.com
------------------------------------------------------------------------
Received on 2003-06-03 17:26:04