基于GPU数据库系统的并发查询性能优化

CONCURRENT QUERY OPTIMIZATIONS BASED ON GPU DATABASE SYSTEM

随着通用图形处理器(GPGPU)计算技术的流行,利用GPU的并行计算能力优化查询执行的性能成为数据库方向的研究热点。现有的研究成果能够利用GPU的高性能计算能力,通过查询任务间协同进行GPU资源管理的机制,支持并发的查询请求,有效提升GPU的资源利用率。但是这种系统架构中由于各查询任务单独管理GPU资源带来重复开销,并且过度使用PCIe总线的数据传输带宽,导致GPU的整体资源利用率仍然较低。Hyper Qx-GPU是一种GPU内存数据库系统新的软件架构设计与实现,该系统通过共享CUDAContext和数据库列存储数据等方式,来提升GPU资源利用率。实验结果表明,相比于当前的GPU数据库系统,Hyper Qx-GPU能够达到平均12.0倍的性能提升。

With the increasing popularity of General-Purpose Graphics Processing Units （GPGPU）, using GPU’s parallel computing power to accelerate database query execution becomes a trend in database research. The state-of-the-art GPU database system can support concurrent queries on GPU by providing a cooperative GPU resources management mechanism, leading to remarkable improvement on GPU resource utilization. However, this kind of system architecture design spawns separate GPU management workload for every single query, which results in duplicated work and limited performance. Furthermore, its excessive use of scarce PCIe bandwidth makes the GPU resource utilization low. This paper presented the design and implementation of a new GPU database system HyperQx-GPU, which improved GPU resource utilization by sharing CUDA Context and database storage. Experiments show that, in comparison with the state-of-the-art GPU database system MultiQx-GPU, the HyperQx-GPU system can achieve 12.0 times of performance improvement on average.