TECHNIQUES FOR DYNAMIC LOAD BALANCING IN SCIENTIFIC SOFTWARE

RESEARCHERS ARE DEVELOPING AND DEMONSTRATING A MODEL FOR SCALABLE HIGH PERFORMANCE COMPUTING SOFTWARE WHICH COMBINES TWO EXISTING COMPUTING PARADIGMS. THE FIRST IS A MODEL FOR DYNAMIC LOAD BALANCING THAT HAS BEEN TESTED ON MULTIPLE-INSTRUCTION MULTIPLE-DATA (MIMD) MACHINES. THE SECOND IS A REUSABLE SOFTWARE DESIGN CALLED A "FRAMEWORK." THE FRAMEWORK IS AN OBJECT-BASED SOFTWARE ARCHITECTURE THAT ASSISTS THE PROGRAMMER IN MODIFYING EXISTING CODES AND IN BUILDING NEW SCIENTIFIC CODES. IT ALSO IMPROVES PROGRAMMER PRODUCTIVITY BY ASSISTING PROGRAM COMPONENT AND DESIGN REUSE. THE RESULTING SOFTWARE IS BEING DEMONSTRATED ON TWO OR MORE FINITE ELEMENT AND FINITE DIFFERENCE NUMERIC CODES. AN EXISTING FRAMEWORK IS BEING ADAPTED FOR THESE PARTIAL DIFFERENTIAL EQUATION CODES TO SUPPORT DYNAMIC LOAD BALANCING AND PROCESSING ELEMENT MIGRATION. THE FRAMEWORK ALREADY PROVIDES MANY OF THE SOFTWARE ARCHITECTURAL FEATURES NECESSARY FOR SUCH A DEVELOPMENT. THESE CODES, BECAUSE OF THE FLEXIBILITY OF THE REPRESENTATIONAL FRAMEWORK, ALSO RUN ON VECTOR OR WORKSTATION PROCESSORS. TOOLS FOR MANIPULATING THE FRAMEWORK ARE BEING ADAPTED TO SUPPORT THE NEW CAPABILITIES AS NECESSARY.