Research On Parallelization Of Scientific Computing Kernels On Multi-core Platform

Accelerating large-scale scientific computing by using heterogeneous multicore/many-core systems could be a future design trend. However, compared with the rapid development of hardware systems, the development of parallel software programming model, especially the models for heterogeneous multicore programming model, is lagging behind. How to translate the potential provided by the hardware into performance and achieve high efficieny in the heterogeneous multicore platform has become of utmost concern to the parallel programming.This thesis presents a parallel computing model for heterogeneous multicore systems, named MS-BSP. Different from the conventional BSP model, it can reflect the major characteristics of heterogeneous multiprocessor system-on-chip (MPSoC), that is, in heterogeneous multicore processors, different kinds of computing tasks can be assigned to different processor kernels in order to provide more flexible and more effective parallel execution, to guide the design and analysis of parallelization scientific computing algorithms on heterogeneous MPSoC. Based on the MS-BSP model, we proposed the a scientific computing parallel programming and optimization framework. Different from the need for explicit management in IBM’s Cell and Nvidia’s CUDA, the mapping of the multi-thread kernel function is managed by the operating system in our framework. As a result, the message-passing based memory access and the synchronization can be greatly reduced and the programming becomes easier. Finally, we implement the parallel programming interface of the operating system which is compatibility with the MPI function to enhance portability.Based on the proposed parallel programming framework, we implement five scientific computing kernels on both the RED and Cell platform and evaluate the performance, which verifies the effectiveness and efficiency of our proposed framework. Since MS-BSP model is deeply optimized on RED platform, the cost of task scheduling is significantly reduced and the efficiency is over 75.67%, while the efficiency is over 63.91% on the Cell platform.