Other Mathlib than Atlas
MKL or ACML Blaslibraries: Both are up to 30% faster than Atlas, but need registering and installing.
[For
Gamess 25 MAR 2010 (R2) only: download
config which has been corrected for bugs and replace the original in ~/gamess] Run
./config composing a new
install.info with the path to your mathlibrary (e.g. /opt/acml4.4.0). Then run lked again with a new name for gamess, e.g.
./lked gamess 01 >& lked1.log. Important for ACML: The download for the 32-bit package is acml-4-4-0-gfortran-32bit.tgz, for the 64-bit target acml-4-4-0-gfortran-64bit-int64.tgz, the library without the "int64" will not work!
Note for linux32 users:
Gamess 25 MAR 2010 (R2): For linking to the MKL libraries lked (the linker script of gamess) has to be edited. Download
lked_mkl32 to replace the original, faulty lked, run ./config and then
./lked_mkl32 gamess 02 >& lked_mkl32.log .This is correct for mkl 10.2.5.035, check this number!
Gamess 1 OCT 2010 (R1): you have to use the original lked and edit as follows:
Line 409: change
case 11help: to
case 11:
Line 412: change
libmkl_intel_lp.a to
libmkl_intel.a
Use the original
config and set GMS_MKL_VERNO to 11 (you may download an edited lked as
lked32 and a correct
install.info, check the mkl release number, e.g. 10.2.5.038).
Rungms on multicore node
rungms has been edited to
make use of up to four cores of a dual or quad core CPU, running in parallel. Just call Gamess with:
rungms job(.inp) 00 2 (or 3, or 4) > job.log to engage the processors of your SMP. With
rungms job(.inp) 00 1 job.out or
rungms job(.inp) > job.log (with gamess.00.x) the normal single CPU run is started. If you use
gameix, then you can select the number of CPU's in the Configuration file on the MasterMenu.
Make sure to check whether Gamess knows how to run your jobtype in parallel. There are many that are not (yet) parallel enabled.