hello all,
i had a couple questions regarding compiling amber 9 on a quad (2
dual core
processor) mac g5 with xlc6.0/xlf8.1 and mpich2 with the latest
patches for
amber 9 installed.
first, on the serial compile, after a few fixes i found on the os x
blog provided
by mengjuei hsieh, everything went well, and in the testing most
things passed,
till a failure at:
cd rdc; ./Run.dip
STOP 1
./Run.dip: Program error
make: *** [test.sander.BASIC] Error 1
if look at rdc/gcg.dip.o i find at the end of the file:
Alignment info will be read from file: RST.dip
Here are comments from the alignment input file:
namelist reports error reading &align
i do have the environment variable:
XLFRTEOPTS=namelist=old:xrf_messages=no
so i don't understand the error. any help would be appreciated.
the second question is regarding multiple processors. i did compile
the parallel version, with mpich2, and it does run, and pass the tests.
using mpd, i can get sander to use all 4 processors, but they don't seem
to get much beyond 70% of each. is there a way to use openmp, or
some other shared memory multiple processor implementation? i
experimented a little with using "xlf_r -qsmp", but couldn't get it to
work. my impression was that the ibm compilers produced faster
code, but i was also thinking about trying the gcc and gfortran
compilers with "funroll-loops -faltivec -Wl,veclib", but here again
i am not sure the altivec vector implementation, even if it can
potentially double or even quadruple the speed, would work
for amber, because of the use of single precision reals. any
comments on these issues would be greatly appreciated. in
addition to the cray and sgi, new cheap shared memory machines
will be available shortly, so though the mpi model has worked
well for clusters, revisiting shared memory models might be
worth it.
thanks again,
--knut
Boston Biomedical Research Institute... Today's Research for
Tomorrow's Health.
Please visit us at www.bbri.org
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun May 13 2007 - 06:07:15 PDT
