[CPMD-list] CPMD for Beowulf Cluster
Christian Tuma
ct at chemie.hu-berlin.de
Fri Aug 23 12:25:38 CEST 2002
Hello Haibo,
you can try
$ ./Configure PC-PGI-MPI > Makefile
You can consider the following points to get improved performance:
(I did some tests with different options to build CPMD 3.5.2, I'm
not totally sure whether my results are true in general)
1. Instead of the PGI math-libs (lapack, blas) use the ATLAS-libs
tuned for your processor type (see ATLAS web pages). Don't forget
to build a complete liblapack.a since the ATLAS-lapack is incomplete
(see ATLAS web pages for instructions as well). Compared to the
PGI-libs the ATLAS-libs give an impressing performance boost (my tests
have shown improved timings of up to 30 % ...)
2. Using the Intel compiler suite (6.0) instead of PGI compilers can
improve the performance only slightly. Since I don't have a Pentium4
at hand I didn't test the option "-tp p7" of the Intel compiler for
Pentium4-processors and therefore cannot compare to the PGI-compiler.
3. MPI-software: On our Linux cluster (Athlon 1800+, 100Mbit ethernet)
the LAM-MPI 6.5.6 gave better wall times compared to MPICH 1.2.4 by
about 15 %. Although pure CPU times (from the CPMD output) are better
with MPICH the communication between the boxes seems to be more effective
using LAM-MPI (although the processor load is still in the range of
0.7 ... 0.8).
Good luck,
Christian.
--
Christian Tuma Humboldt-Universitaet Berlin
ct at chemie.hu-berlin.de Arbeitsgruppe Quantenchemie (Prof. Sauer)
phone: +49-30-20937140 Brook-Taylor-Str. 2, 12489 Berlin, GERMANY
fax: +49-30-20937136 http://www.chemie.hu-berlin.de/ag_sauer
More information about the CPMD-list
mailing list