作　　者： ; ; ; ; ;

机构地区： 华南理工大学计算机科学与工程学院

出　　处： 《华南理工大学学报（自然科学版）》 2013年第3期22-28,共7页

摘　　要： 为进一步提高内存数据库索引结构T-树的操作性能,提出一种基于图形处理器的T-树无锁并行计算方案.该方案通过分析平衡树结构的父子节点间的关系,在图形处理器平台上实现使用m个线程并行创建具有m个节点的T-树索引,从而以最大并行度的方式构建T-树.为验证方案的正确性,提出以堆栈的方式在图形处理器上遍历T-树的算法,对各平台上构建T-树的方案进行性能分析,并通过页锁定内存的方式提高CPU和GPU间的数据传输速率.通过对多个处理器平台上的实验结果的对比发现,提出的方案在并行构建T-树和T-树的批量节点插入上相比于传统CPU平台方案分别获得12倍和8倍以上的加速比. In order to improve the operation performance of T-tree, a main memory database indexing, a lock-free parallel computing scheme for T-tree on the graphics processing unit （GPU） is proposed. By analyzing the parent- child relationship of balanced tree structure, a T-tree with m nodes is constructed with n separate threads in a parallel mode on GPG platform. Thus, a T-tree with the maximum parallelism degree is successfully constructed. Then, in or- der to verify the correctness of the proposed scheme, a T-tree traversal algorithm with stacking on GPU platform is proposed. Moreover, different schemes of constructing T-tree on different platforms are analyzed and the data trans- mission between CPU and GPU is optimized by means of memory pinning. The result of experiments on different plat- forms show that, as compared with the conventional CPU-based schemes, the proposed scheme obtains about 12 times batch creating and 8 times batch inserting of T-tree.

关 键 词： 图形处理器 树 内存数据库 索引结构 并行构建 批量节点插入

领　　域： [自动化与计算机技术] [自动化与计算机技术]