应用动态生成树的GPU显存数据复用优化  

Optimization Method for Reusing GPU Data Based on Dynamic Spanning Tree

作　　者：李亮[1] 王恩东 朱正东[1] 颜康[1] 张保[1] 董小社[1] 

机构地区：[1]西安交通大学电子与信息工程学院,西安710049 [2]浪潮集团高效能服务器和存储技术国家重点实验室,济南250013

出　　处：《西安交通大学学报》2013年第10期44-50,共7页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61173039);国家"863计划"资助项目(2012AA010904;2012AA01A306);国家科技支撑计划资助项目(2011BAH04B03)

摘　　要：针对手工优化GPU（Graphic Processing Unit）显存级数据复用过程复杂和编译时优化数据复用开销过大的问题,提出了一种基于动态生成树在运行时进行数据复用的优化方法,可为程序员提供一种透明且高效简单的优化方式.该方法将已经执行的GPU计算任务的数据访问抽象为生成树的叶子节点,利用动态生成树管理GPU数据访问信息,实现了运行时GPU显存级数据的复用优化,并通过运行时对生成树的搜索和维护,动态地发掘和优化GPU显存级数据复用,因此,在编程时不需要进行复杂且困难的数据复用分析,直接调用文中提出的运行时库就能有效减少程序执行过程中CPU内存和GPU显存之间的冗余数据传输次数,从而提升应用的运行性能.实验结果表明,使用文中提出的优化方法可有效消除未进行数据复用优化的CPU-GPU应用程序中的冗余数据传输,最大加速比达原始执行的3～10倍,额外开销不到优化后程序总执行时间的5％.An optimization method for reusing graphic processing unit （GPU） data is proposed based on the dynamic spanning tree to solve the problem that the reusing process with manual optimization is complicated and compilation overhead in optimization is expensive.The proposed method is transparent to programmers with simplicity and effectiveness.The approach abstracts the executed accesses of GPU tasks as leaf nodes of a spanning tree,and uses the tree to dynamically manage the access information.Then the identification and optimization for the data reusing is realized by searching and managing the tree.The complex and challenging data reuse analysis is not required since invoking the run-time library is sufficient to reduce the CPU-GPU data transfers.Experiments show that the proposed optimization method can eliminate redundant CPU-GPU data transfers in non-reuse CPU-GPU application programs and can achieve a speedup as large as 3 to 10 times over the original execution.Moreover,the additional cost is just less than 5 ％ of the execution time of the program adopting the proposed optimization.

关 键 词：GPU显存 动态生成树 数据复用 数据传输 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]