A general parallelization approach to improve computation efficiency in a global chemical transport model (GEOS-Chem)

抄録

Despite increasing processing power of systems, sequential programs have limitations to achieve corresponding performance gains. Parallelization of sequential programs can exploit the computation potential to reduce overall execution expenditures. However, parallelization of a very large legacy program arises great challenges, because 1) former design of the program did not consider parallelization, 2) the large-scale program has been developed for long time even decades, and 3) some developers of the program are no longer available. In the atmospheric science field, there are many such legacy programs and GEOS-Chem model developed in Harvard University is one of examples. In this paper, we suggest simple procedures of the MPI parallelization which can be easily applied to a large legacy program. We applied this method to the GEOS-Chem model which is already parallelized with the OpenMP method alone. Our test results showed that the good performance improvement can be obtained by parallelizing a small portion of the program.