异构体系结构下的混合负载数据库自适应并行技术

Adaptive parallelism technology for mixed-load database in heterogeneous architecture

混合负载处理(HTAP)数据库通过统一的平台处理不同的工作负载,消除数据在不同平台之间迁移和开发运营的成本,实现对最新事务数据的实时分析。针对在保证系统新鲜度的同时最小化在线事务处理(OLTP)和在线分析处理(OLAP)在系统运行过程中的相互干扰的问题,提出一种中央处理器/图形处理器(CPU/GPU)异构体系结构下的混合负载数据库自适应并行技术。首先,构建异构体系结构下的混合负载数据库系统;其次,针对系统中的数据新鲜度问题,提出基于固定时间段的任务调度方法,以分离数据同步与分析查询;再次,提出一种多流并行分析处理策略,以提高系统分析处理性能;最后,针对高新鲜度的场景提出热数据优化方法,以提供全局最高新鲜度。基于CH-benCHmark基准进行测试和对比实验结果表明,与BatchDB相比,所提混合负载数据库事务性能下降率减小了50%。所提优化策略可以在提供可预测的新鲜度的情况下提供最高的分析处理性能,在少量性能损失的情况下提供最高的新鲜度。

Hybrid Transactional Analytical Processing(HTAP)databases handle different workloads on a unified platform,eliminating the cost of data migration between different platforms and enabling real-time analysis of the latest transaction data.To address the issue of minimizing the mutual interference between OnLine Transaction Processing(OLTP)and OnLine Analytical Processing(OLAP)while ensuring system freshness,an adaptive parallel technique for hybrid workload database was proposed under the heterogeneous Central Processing Unit/Graphics Processing Unit(CPU/GPU)architecture.Firstly,a hybrid workload database system under the heterogeneous architecture was constructed.Then,a task scheduling method based on fixed time periods was proposed to separate data synchronization from analytical queries to address the system’s data freshness issue.Furthermore,a multi-stream parallel analysis processing strategy was suggested to improve system analytical processing performance.Finally,a hot data optimization method was put forward for high freshness scenarios to provide the highest global freshness.Experiments based on the CH-benCHmark show that,compared with BatchDB,the proposed hybrid workload database achieves a 50%reduction in transaction performance degradation.The proposed optimization strategy can provide the highest analytical processing performance while offering predictable freshness,as well as the highest freshness with minimal performance loss.