Parallel execution of a Fortran 77 weather prediction model

This work presents the first MPP execution of the ARPS storm prediction code, und it is a significant step beyond the best previous runs on the fustest machines uva~lable. ARPS, a real third-party application developed in Fortran 77 by scientists for their own use, was greatly sped up though execution on a massively parallel machine. There are three notable features to our erLtry. First, the execution was made possible because of new automatic parallelizing compiler technology, a new tool called CMAX. The entire 65,000 line program was translated, compiled, and then executed on a CM-5; our results were achieved with the help of G’MAX rather than brute force rewriting. Second, we achieved good speedups as we increased the number of processors. Finally, performance on large CM-5’S is sign~ficantly better than performance on any previous ARPS platform. We obtained the following results. First, a speedup of 907 on a 1024 processor CM-5. The best previous speedup of ARPS on a parallel platform was 8..2. Second, that run was 57 times faster than the fastest previously available ARPS platform with published timings, a Cray Y-MP. Finally, we achieved at least 2 times the price/performance of a Cray Y-MP.