Many of us are only interested in applications rather than the science of computation although they are not entirely unrelated. If I had say a program written in standard MPI C, (using say the OpenMPI software), can I get this program to run on CUDA without going into the program source code to tweak it into a form that allows for the MPI CUDA compilation?
In particular, here we are running the MPI version of Espresso MD package (http://espressowiki.mpip-mainz.mpg.de/wiki/index.php/Main_Page) where the program is written in C, but is driven by a TCl script. You have to specify when you compile the program whether you are running mpi or not. How does one integrate this program with CUDA without in any way changing the internal source files so that it can run on CUDA MPI?
Any help in helping us use the above package with CUDA would be greatly appreciated.
Also, how many MPI runs of a program (separate jobs) can be executed simultaneously? Without a Scheduler, it must be a tricky problem for the computer system.
Please use clear, simple step by step language without presupposing any high level competence of UNIX, setups or even how they are expected to be integrated since we are raw beginners.