一种基于GPU的移动对象并行处理框架  

A GPU-BASED MOVING OBJECT PARALLEL PROCESSING FRAMEWORK

作　　者：韦春丹 龚奕利[1] 李文海[1,2] 

机构地区：[1]武汉大学计算机学院,湖北武汉430072 [2]软件工程国家重点实验室,湖北武汉430072

出　　处：《计算机应用与软件》2016年第10期223-229,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61100020);华为公司创新研究计划资助项目

摘　　要：PGrid是一个基于格网索引的移动对象并行处理框架。通过分析PGrid框架不利于在GPU上并行的因素,提出基于GPU的无锁并行处理G-LFPP(GPU Based Lock Free Parallel Processing)框架。采用基于操作分解/聚类的无锁更新策略,消除更新过程中并发控制对更新性能的影响;为了实现细粒度并行查询,提出基于候选集映射表和查询确认表的快速查询索引。实验表明,该方法更新和查询策略有利于大规模线程并发处理更新和查询。当移动对象的数量达到千万级时,更新速率和查询速率仍然可以超过每秒1100万次和110万次。与PGrid相比,并发处理更新和查询的速度提高了6.61倍。PGrid is a parallel processing framework for moving objects based on grid index. This paper proposes a GPU-based lock-free parallel processing( G-LFPP) framework on the basis of analysing the factors of PGrid framework not being conducive to parallel on GPU. It uses a lock-free update strategy,which is based on operation splitting/clustering,to eliminate the impact of concurrency control on update performance in updating process. In order to implement fine-grained parallel queries,the paper also presents a fast query index composed of candidate set mapping table and query confirmation table. Experiments show that the proposed update and query strategies avail the large-scale threads in processing updates and queries concurrently. Whilst the number of moving objects reaches a level of ten million,the update and query rates can still be over 11 million per second and 1. 1 million per second respectively,which are 6. 61 times higher than those of PGrid.

关 键 词：并行计算 图形处理单元 异构计算 格网索引 移动对象数据库 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]