指令感知的nvmeof存储网络报文处理机制
首发时间:2021-10-25
摘要:nvme over tcp可基于现有数据中心网络基础设施和标准的tcp/ip软件协议栈实现nvmeof(nvme over fabrics)存储网络,是未来数据中心构建大规模高性能存储系统的关键技术。本文设计实现指令感知的nvmeof报文处理机制panot(pdu-aware nvme over tcp),首先通过基于聚合pdu的中断合并技术缓解中断风暴问题并实现高吞吐率,其次通过r2t指令感知的pdu发送机制缓解指令依赖性问题并有效降低写请求延迟。本文在linux内核中实现了panot原型系统,性能测试结果表明,相比标准linux内核中的nvme over tcp实现,panot可将cpu资源消耗减少50%以上,将iops最多提升48.86%。
关键词: nvme over fabrics nvme over tcp 存储网络 中断合并 指令感知
for information in english, please click here
pdu-aware transmission mechanism for nvmeof storage networks
abstract:nvme over tcp is adequate for building large-scale high-performance nvmeof (nvme over fabrics) storage networks for future data centers using existing network infrastructure and tcp/ip software stacks.in this paper, we propose panot, a pdu-aware nvme over tcp storage network implementation. panot introduces pdu groups to aggregate pdus before transmission to alleviate the interruption storm problem and achieve high throughput, proposes r2t pdu-aware transmission mechanism to alleviate the dependency issues of h2cdata to r2t and reduce write latency. panot prototype was implemented in linux kernel and compared with standard nvme over tcp implementation. performance evaluation results show that panot can reduce cpu utilization by more than 50% and improve iops by up to 48.86%.
keywords: nvme over fabrics, nvme over tcp, storage networks, interrupt coalescing, pdu-aware
论文图表:
引用
导出参考文献
no.****
同行评议
共计0人参与
勘误表
指令感知的nvmeof存储网络报文处理机制
评论
全部评论