Optimizing load balance and communication on parallel computers with distributed shared memory

摘要

To optimize programs for parallel computers with distributed shared memory two main problems need to be solved: load balance between the processors and minimization of interprocessor communication. This article describes a new technique called data-driven scheduling which can be used on sequentially iterated program regions on parallel computers with a distributed shared memory. During the first execution of the program region, statistical data on execution times of tasks and memory access behaviour are gathered. Based on this data, a special graph is generated to which graph partitioning techniques are applied. The resulting partitioning is stored in a template which is used in subsequent executions of the program region to efficiently schedule the parallel tasks of that region. Data-driven scheduling is integrated into the SVM-Fortran compiler. Performance results are shown for the Intel Paragon XP/S with the DSM-extension ASVM and for the SGI Origin2000.