Common block in Domain Decomposition implementation
by Yi Zhang
Fortran 77 essentially does not have global variables for subroutines, instead, common blocks are used to passing data across subroutines when arguments are not preferred. This is conceptually unsafe, though widely used in some legacy F77 codes. The code I’ve been working on adopts this “feature” extensively. The problem is, in Dirichlet-Neumann-like domain decomposition implementation, the solver will be called twice for each iteration, each time with different stiffness matrix setup(due to different subdomains) and boundary conditions. On the other hand, the common blocks make those two calls undistinguishable.
To reuse the legacy solver, my first try was to construct two copies of the solver by putting it into two modules, hoping the explicit interface of the modules would identify the common blocks in different copies. This turns out to be wrong: the subroutines from the two modules still share the same common blocks.
What I finally decided to do is to build a copy of the original solver by renaming all the common blocks. Sed & AWK make this not to difficult. But still I don’t think that’s the most efficient way to tackle this. Consider this is a lesson that not all sequential solvers are easy to parallelize.