Software issues on mapping applications onto heterogeneous machines and the performance of Krylov algorithms on parallel machines
Abstract
One way to exploit a mixed-machine heterogeneous computing environment is to decompose an application task into subtasks, assign each subtask to a machine (matching) and order the execution (scheduling) such that the overall task execution time is minimized. Another way is to consider a meta-task consisting of a collection of independent tasks, i.e., with no data dependencies or precedence constraints among the tasks. For such cases, the mapping problem is the minimization of the completion time of the overall meta-task consisting of all the tasks. This thesis develops dynamic heuristics for performing matching and scheduling (mapping) for both cases. First, novel schemes are developed to use information that becomes available at run time to improve a statically obtained initial mapping. As part of these schemes a dynamic heuristic called the hybrid remapper is designed. The hybrid remapper bases its decisions on a mixture of run-time and expected values for the subtask completion times and machine availability times. The potential of the hybrid remapper to improve the performance of initial static mappings is demonstrated using simulation studies. Second, dynamic heuristics for mapping meta-tasks onto heterogeneous computing systems is presented. The heuristics can be grouped into two categories: on-line mode and batch mode. Simulations are performed to compare the performance of the different heuristics. Performance and scalability of the conjugate gradient squared (CGS) algorithm on parallel machines are studied. Factors that contribute toward the synchronization and communication overheads are examined. Based on the results, a modified CGS (MCGS) algorithm is proposed. Furthermore, performance and scalability of several preconditioners are studied both experimentally and theoretically on machines with significant inter-processor communication times.
Degree
Ph.D.
Advisors
Siegel, Purdue University.
Subject Area
Computer science|Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.