ContentsIndex

3.8.3 Hewlett-Packard: V2500

The 12-processor machine ``zeipel'' fromHewlett-Packard is a ``V2500 PA 2.0'' system. Now, there is a first success to force the compiler to accept the OpenMP directives in CO5BOLD. Yet, when running on several processors, only some routines (e.g. rhd_shortrad_dirsimple1) in CO5BOLD can benefit while others (rhd_shortrad_dirsimple2, rhd_shortrad_dirsimple3) are significantly slower than on one processor. In addition, the single-processor performance is not very good, partly because the achievable optimization level is not very high.

Some macros, which seem to be necessary:

• +U77: Link proper library to make the machine understand e.g. call flush(6).
• +cpp=yes: Switch on the C preprocessor. Note that all Fortran90 files have to end with ``.f90''. The ``.F90'' suffix does not seem to work.
• +Oparallel +Oopenmp +Onoautopar: Try to enable parallelization with OpenMP directives, disable auto-parallelization.
• +Onoinline: Disables inlining. This can simplify things. With a proper choice of routine versions inlining is not really necessary anymore.
• +O3 +Olimit: General optimization with limited resource usage during compilation. Some modules should only be compiled with +O2, others compile even with +O3 +Onolimit.

 

 

The UIO modules and the string handling module should be compiled in debug mode. A proposed compiling sequence is (MSrad does not compile; all other modules are activated):
 export F90_LHDRAD=1
export F90_MSRAD=0
export F90_SHORTRAD=1
export F90_DUST=1
export F90_MHD=1
export F90_PARALLEL=openmp

export F90_DEBUG=1
./configure
make UIO STR

export F90_DEBUG=0
./configure
make