基于MPI的卷积计算并行实现

Implementation of Parallel Convolution Based on MPI

摘要:

针对传统的卷积并行计算模型中,存在着大量的消息传递,负载不均衡等问题;提出一种新的基于MPI同步模型的并行卷积算法;该模型采用消息传递的方式进行进程间的通信,同时有效平衡负载,避免大量的消息传递;通过分析该模型的加速比和效率,实验结果表明,此方法显著提高了并行效率和长序列的运算速度,充分发挥了节点间分布式存储和多核并行处理的优势,是一种有效可行的并行策略。

Abstract:

Convolution computing plays an important role in scientific computing. There exist massive message passing, load-unbalancing in traditional MPI model. Concerning this question, a new parallel convolution based on MPI model was proposed in this paper. This model prevents massive message passing, and keeps load-balancing. By analyzing the speedup ratio and efficiency, the experimental result shows that the new MPI parallel programming obviously promotes parallel efficiency and large-scale sequence’s speed, which exerts the advantages of distributional memory between nodes and parallel computing. It is an effective and feasible parallel strategy.