摘要
新兴的持久性内存技术的出现和RDMA网络为构建新型的远程空间数据索引提供了新的可能,有望为R树这样的空间数据结构带来性能提升.由于持久性内存相比DRAM具有更慢的访问时延、额外的持久化开销以及缺少远程持久化原语等因素导致完全持久化的远程R树性能下降.在此基础上,本文设计并实现了RRtree,一种基于混合部署架构的远程持久性R树,在保证数据持久性的前提下实现尽可能高的性能.同时,通过选择性元数据持久化、写合并和对双边RDMA原语结合持久性内存使用进行优化.最后,在真实的傲腾持久性内存上实现并验证了RRtree的高性能和高可扩展性.实验结果表明,RRtree和对比对象FBR-tree相比其时延和吞吐分别有显著的降低和提升.
The emergence of burgeoning persistent memory technologies and RDMA networks offer new possibilities for building new types of remote spatial data indexes,promising performance gains for spatial data structures like R-trees.The performance of purely persistent remote R-tree is degraded due to the slower access latency of persistent memory compared to DRAM,the extra persistence overhead,and the lack of remote persistence primitives.Based on this,this paper designs and implements RRtree,a remote persistent R-tree based on a hybrid deployment architecture,to achieve as high performance as possible while guaranteeing data persistence.In addition,it is optimized by selective metadata persistence,write-coalescing,and combination with persistent memory throughtwo-side RDMA primitives.Finally,the high performance and scalability of RRtree is implemented and verified on real Optane persistent memory.The experimental results show that RRtree shows significant reduction and improvement in latency and throughput,respectively,compared to the comparison object FBR-tree.
作者
吴瑶
张瑞
吴杰
WU Yao;ZHANG Rui;WU Jie(School of Computer Science,Fudan University,Shanghai 200433,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第12期2858-2864,共7页
Journal of Chinese Computer Systems
基金
国家重点研发计划课题项目(2021YFC3300600)资助。
