A Programming Framework Based on Multi-GPU

Gang WANG,Jie TANG,Gang-shan WU
DOI: https://doi.org/10.3969/j.issn.1673-629X.2014.01.003
2014-01-01
Abstract: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.
What problem does this paper attempt to address?