Recent advances in PC-Linux systems for electronic structure computations by optimized compilers and numerical libraries

Jen-Shiang Yu*, Chin Hui Yu

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

7 Scopus citations

Abstract

One of the most frequently used packages for electronic structure research, GAUSSIAN 98, is compiled on Linux systems with various hardware configurations, including AMD Athlon (with the "Thunderbird" core), AthlonMP, and AthlonXP (with the "Palomino" core) systems as well as the Intel Pentium 4 (with the "Willamette" core) machines. The default PGI FORTRAN compiler (pgf77) and the Intel FORTRAN compiler (ifc) are respectively employed with different architectural optimization options to compile GAUSSIAN 98 and test the performance improvement. In addition to the BLAS library included in revision A. 11 of this package, the Automatically Tuned Linear Algebra Software (ATLAS) library is linked against the binary executables to improve the performance. Various Hartree-Fock, density-functional theories, and the MP2 calculations are done for benchmarking purposes. It is found that the combination of ifc with ATLAS library gives the best performance for GAUSSIAN 98 on all of these PC-Linux computers, including AMD and Intel CPUs. Even on AMD systems, the Intel FORTRAN compiler invariably produces binaries with better performance than pgf77. The enhancement provided by the ATLAS library is more significant for postHartree-Fock calculations. The performance on one single CPU is potentially as good as that on an Alpha 21264A workstation or an SGI supercomputer. The floating-point marks by SpecFP2000 have similar trends to the results of GAUSSIAN 98 package.

Original languageEnglish
Pages (from-to)673-681
Number of pages9
JournalJournal of Chemical Information and Computer Sciences
Volume42
Issue number3
DOIs
StatePublished - 1 May 2002

Fingerprint Dive into the research topics of 'Recent advances in PC-Linux systems for electronic structure computations by optimized compilers and numerical libraries'. Together they form a unique fingerprint.

Cite this