PROGRAM MAPPING STRATEGIES FOR MULTIPROCESSOR COMPUTERS

THE PARADIGM COMPILER IMPLEMENTS A NEW TECHNOLOGY FOR COMPILING PROGRAMS FOR SCIENTIFIC COMPUTATION FOR EXECUTION BY LARGE-SCALE MULTIPROCESSOR COMPUTERS. THE TECHNOLOGY IS BASED ON AN ANALYSIS OF THE SOURCE PROGRAM INTO CODE BLOCKS,EACH OF WHICH DEFINES A DATA STRUCTURE. EACH SUCH DATA STRUCTURE IS MAPPED INTO THE COLLECTION OF PROCESSING ELEMENTS OF THE TARGET COMPUTER SO THAT THE COLLECTION OF MAPPINGS YIELDS THE BEST POSSIBLE PERFORMANCE. THE PROPOSEDEFFORT WILL: (1) PROVIDE A USER-FRIENDLY INTERFACE THROUGH WHICH USERS MAY ADVISE THE COMPILER REGARDING MAPPING CHOICES FOR EACH CODE BLOCK; (2) IMPLEMENT MEANS FOR ESTIMATING PERFORMANCE OF THE COMPILED CODE AS A FUNCTION OF TARGET MACHINE PARAMETERS AND MAPPING DECISIONS; (3) EVALUATE VARIOUS MAPPING STRATEGIES FOR SEVERAL SIGNIFICANT APPLICATION CODES; AND (4) DEMONSTRATE THE COMPILING TECHNOLOGY AT CONNECTION MACHINE USER SITES. THE BENEFITS OF THIS PROJECT WILL BE: EASIER DEVELOPMENT OF EFFICIENT CODES FOR MULTIPROCESSOR COMPUTERS; THE CLARITY OF EXPRESSION FOR SCIENTIFIC COMPUTATION MADE POSSIBLE BY USE OF A FUNCTIONAL PROGRAMMING LANGUAGE; AND THE TRANSFER OF NEW TECHNOLOGY INTO COMPUTING PRACTICE. A SPECIFIC AND UNIQUE BENEFIT IS THAT THE USER OF PARADIGM CAN EXPERIMENT WITH SEVERAL MAPPING CHOICES WITHOUT HAVING TO MAKE ANY CHANGES TO THE HIGH-LEVEL SOURCE LANGUAGE PROGRAM.