异构计算平台上列存储系统的并行连接优化策略

发布时间：2018-11-24 14:13

【摘要】：GPU以及集成式的CPU-GPU架构凭借其强大的并行处理能力和可编程流水线方式,已经成为数据库领域的研究热点。为充分利用异构平台的并行计算能力,提升列存储系统的查询性能,在研究异构平台结构特性的基础上,首先提出了GPU多线程平台上进行连接的数据划分策略——ICMD(Improved CMD),利用GPU流处理器并行处理各个子空间上的连接,然后利用任务评估分配模型实现查询负载的动态分配,使得查询操作能在多核CPU、GPU上高效并行执行。同时利用片上全局同步机制、局部内存重用技术优化ICMD连接算法。最后采用SSB基准测试集测试,结果表明:Intel~?HD Graphics 4600平台上并行连接查询相比于CPU版本获得了35%的性能提升,较GPU查询引擎的Ocelot性能上提升了18%。
[Abstract]:With its powerful parallel processing ability and programmable pipeline, GPU and integrated CPU-GPU architecture have become the research hotspot in database field. In order to make full use of the parallel computing ability of heterogeneous platform and improve the query performance of column storage system, on the basis of studying the structural characteristics of heterogeneous platform, this paper first puts forward the data partition strategy of joining on GPU multithreaded platform, which is called ICMD (Improved CMD),. The GPU stream processor is used to process the connections in each subspace in parallel, and then the task assessment allocation model is used to realize the dynamic allocation of query load, which enables the query operations to be executed in parallel on multi-core CPU,GPU efficiently. At the same time, the ICMD connection algorithm is optimized by using the on chip global synchronization mechanism and the local memory reuse technology. Finally, using SSB benchmark set test, the results show that the parallel join query on the Intel~?HD Graphics 4600 platform has achieved 35% performance improvement compared with the CPU version, and 18% higher than the Ocelot performance of the GPU query engine.
【作者单位】：东华大学计算机学院;
【基金】：国家自然科学基金(No.61103046) 上海市自然科学基金(No.11ZR1401200)
【分类号】：TP311.13;TP333

【相似文献】