Compiling Fortran D for MIMD distributed-memory machines

for MIMD Distributed-Memory Machines :llel computing resents the only :usible way to continue to increase the computational power available to scientists and engineers. Parallel computers , however, are not likely to be widely successful until they are easy to program. A major component in the success of vector supercomputers is the ability of scientists to write Fortran programs in a "'vec-torizable'" style and expect vectorizing compilers to automatically produce efficient code [8, 35]. The resulting programs are easily maintained , debugged, and ported across different vector machines. Compare this with the current situation for programming parallel machines. Scientists wishing to use such machines must rewrite their programs in an extension of For-tran that explicitly reflects the architecture of the underlying machine. Multiple-instruction, multiple data (MIMD) shared-memory machines such as the Cray Research Y-MP C90 are programmed with explicit synchronization and parallel loops found in Parallel Computing Forum (PCF) Fortran [24]. (SIMD) machines such as the Thinking Machines CM-2 are programmed using parallel array constructs found in Fortran 90 [3]. MIMD distributed-memory machines such as the Intel Paragon provide the most difficult programming model. Users must write message passing Fortran 77 programs that deal with separate address spaces, synchronizing processors, and communicating data using messages. The process is time-consuming, tedious, and error-prone. Significant increases in source code size are not only common but expected. Because parallel programs are extremely machine-specific, scientists are discouraged from utilizing parallel machines because they risk losing their investment whenever the program changes or a new architecture arrives. We propose to solve this problem by developing the compiler technology needed to establish a machine-independent programming model. It must be easy to use, yet perform with acceptable efficiency on different parallel architectures, at least for data-parallel scientific codes. The question is whether any existing Fortran dialect suffices. Parallel Computing Fortran is undesirable because it is easy to inadvertently write programs with data races that produce indeterminate results. Message-passing For-tran 77 is portable, but difficult to use. Fortran 90 is promising, but may not be sufficiently flexible. What all these languages lack is a way to specify the decomposition and placement of data in the program. We find that selecting a data decomposition is one of the most important intellectual steps in developing data-parallel scientific codes. Though many techniques have been developed for automatic data decomposition, we feel that the compiler will not be able to choose an efficient …