MULTI-COMPILER FOR AEROSPACE PARALLEL PROCESSING

IN SEARCH OF COMPUTATIONAL ADVANCES FOR AEROSPACE APPLICATIONS, NASA NEEDS METHODS AND SOFTWARE FOR EASING THETASK OF DEVELOPING EFFICIENT PROGRAMS FOR PARALLEL COMPUTERSWHILE RETAINING EFFICIENCY OF PROGRAMS TRANSFERRED BETWEEN ARCHITECTURAL DESIGNS. THE DESIGN AND OPTIMIZATION OF EFFICIENT AND RELIABLE PARALLEL SOFTWARE REQUIRES A MUCH HIGHER LEVEL OF COMPUTER ASSISTANCE. MOST CURRENT PARALLELIZATION EFFORTS ATTEMPT TOAUTOMATICALLY EXTRACT PARALLEL DESIGNS FROM SCALAR CODE. THESE TOOLS ARE VERY ARCHITECTURE-SPECIFIC, AND ONLY ACHIEVELIMITED SPEEDUP BECAUSE THE DEEPER KNOWLEDGE REQUIRED FOR HIGHER SPEEDUP IS MISSING FROM THE SCALAR DESIGN. THE PROPOSED MULTI-COMPILER APPROACH IS INNOVATIVE BECAUSE IT CONCENTRATES ON THE REVERSE EXPANSION PROCESS -- FINDING EFFICIENT MAPPINGS FROM A SINGLE VERY-HIGH-LEVEL PARALLEL SPECIFICATION INTO ONE OR MORE TARGET PARALLEL ARCHITECTURES. IT UTILIZES A UNIQUE SPECIFICATION MODEL THAT CAPTURES ALL OF THE DIMENSIONS OF A PROBLEM THAT HAVE THE POTENTIAL TO BE MAPPED ONTO THE PROCESSOR TOPOLOGY OF A GIVEN ARCHITECTURE. DURING PHASE I, A PROTOTYPE OF THE MULTI-COMPILER WILL BE DEVELOPED AND EVALUATED TO ESTABLISH THE FEASIBILITY OF THE APPROACH. THE PROTOTYPE WILL BE EVALUATED FOR ITS UTILITY IN DEVELOPING PARALLEL SOFTWARE, THE EFFICIENCY OF GENERATEDTARGET MAPPINGS, AND THE FEASIBILITY OF IMPLEMENTING A FULL SYSTEM.