集成CPU-GPU架构上的列存储连接优化技术研究

集成多核CPU-GPU架构已经成为计算机处理器芯片的发展方向。利用这种架构的并行计算能力进行数据处理已经成为了数据库领域的研究热点。为了提高列存储系统的查询性能,首先改进了已有协处理机制中的负载分配策略,通过监测数据库系统CPU占用率,动态地为处理器提供合理的数据划分；然后,针对集成多核CPU-GPU架构上的数据预取机制,提出了一种确定预取数据大小的模型,同时,针对GPU访存的特点,进行了GPU访存优化；最后,使用OpenCL作为编程语言,实现了一种集成多核CPU-GPU架构上的列存储排序归并连接算法,并采用提出的方法对连接处理进行优化。实验证明,所提优化策略可以使列存储系统排序归并连接性能提升33%。

Column-stores Join Optimization on Coupled CPU-GPU Architecture

Heterogeneous architecture is the new trend of the development of computer system central processor unit(CPU).Taking advantage of its powerful computer power has been a new research hotspot in database system field.First,in order to enhance the query performance of column-oriented database,we proposed a data partition model which is environmentally sensitive.The data partition model provides optimal data division for every processing unit dynamically by monitoring the CPU occupancy rate.Then,for GPU memory access optimization,we proposed a DFAT estimate model for prefetching.At the same time,we optimized GPU memory access based on coalesced access.We implemented a sort-merge join algorithm on a PC with an integrated CPU-GPU chip,which adopts the out data partition model and our cost model in prefetching.Our strategy is able to distribute data to different processing units automatically,and can make sort-merge join achieve a performance improvement of 33% on coupled CPU-GPU architecture.