基于多GPU集群的编程框架

摘要

现如今,GPU作为一种低功耗高性能图形处理器单元,被广泛应用于高度并行化的应用程序中。其线程和内存的层次结构在诸多成功的多线程应用和科学研究中表现出巨大的优势。为了简化多GPU集群的编程模式以及更好地利用GPU的计算性能,设计并实现了一个新的基于多GPU的MapReduce并行编程框架。使用了并行虚拟文件系统( PVFS)来存储数据,考虑了动态的负载平衡和GPU相关的权重要素以达到优化系统的效率、透明性以及系统的可伸缩性的目的。在文中,将演示使用该编程模式解决地质应用的一个典型的偏移应用-叠前时间偏移( PKTM),并给出实验结果。%Nowadays,GPU as a power-efficient high performance unit, is widely used in highly parallel application. Its hierarchy of threads and memory has proven successful at programming multi-thread applications and scientific research. To simplify programming model and better utilize the GPU,design and implement a new parallel programming framework,based on MapReduce. In the project,em-ploy a Parallel Virtual File System ( PVFS) to store data distributely. As an attempt to improve the efficiency,transparent and scalability, take the dynamic load balancing and GPU weight factor in cluster into consideration. Demonstrate how typical tasks in oil industry are modified to fit into the programming model. Evaluate programming model using Prestack Kirchhoff Time Migration ( PKTM) applica-tion,and present the experiment results.