GPU集群中的每一个节点配备GPU 卡，使机器的处理能力大幅提高。GPU 集群成为将来集群配置的重要模式，在未来的数据中心成为重要角色。FFT 算法是云计算、数据挖掘、科学与工程计算领域的基础算法。由于GPU 集群存储层次深，大规模FFT 算法在保持高性能的前提下移植到其他机器极具挑战。本项目针对大规模数据集，研究建立可移植的FFT 数学库原型。主要研究内容包括：（1）针对FFT数学库原型这一最终目标，研究GPU集群编程新机制；（2）基于GPU集群编程新机制，研究GPU集群性能模型，开发FFT数学库原型；（3）此研究过程中积累的关于GPU集群编程新机制的技术、方法和工具可以广泛应用于解决其他算法的GPU集群移植问题。项目研究成果将提高GPU集群硬件平台的使用效率。

In GPU clusters, each computer node is equipped with one or more GPU cards. GPU clusters are becoming more and more popular nowadays and will be widely used in data centers. FFT is a basic algorithm widely used in cloud computing, data mining, scientific and engineering computing. Because of the deep storage hierarchy of GPU clusters, it is a great challenge to transplant a large-scale FFT code from one GPU cluster to another while keeping the high performance. This research seeks to develop a FFT math library prototype of GPU clusters. The research work includes: (1) in order to get a FFT math library prototype, the programming methodology of GPU clusters must be researched; (2) based on the new programming methodology, a FFT math library prototype is developed; (3) the new technology, method and tools applied in the development of FFT math library prototype can be used in future work when transporting other similar algorithms. This research result will improve the using efficiency of GPU clusters.