BOOM-KV:基于RDMA的高性能NVM键值数据库

BOOM-KV:an RDMA-enabled high-performance NVM key-value store

随着英特尔傲腾数据中心持久化内存模块(DCPMM)开始进入市场以及远程直接内存访问(RDMA)硬件成本的降低,设计融合非易失性内存(NVM)和RDMA的键值(KV)数据库面临新的机遇和挑战。构建基于NVM和RDMA的KV数据库的关键在于设计一个高效的通信协议。遗憾的是,现有工作或采用NVM不感知的RDMA协议,或采用低效的NVM感知的RDMA协议,这导致它们无法最大化KV数据库的性能。本文提出了BOOM协议——一种新型的NVM感知的RDMA协议。相较于NVM不感知的协议,BOOM协议允许直接对远端NVM进行RDMA操作,消除了冗余的数据拷贝;相较于现有的NVM感知的协议,它可以显著减少元数据请求,降低KV请求的端对端延迟。在BOOM协议的基础上构建了BOOM-KV,并针对服务端中央处理器(CPU)利用率和宕机持久化等问题进一步进行优化。将BOOM-KV与最新的研究成果进行对比,结果表明,BOOM-KV能显著降低请求延迟,其中PUT延迟最大降低了42%,GET延迟最大降低了41%,并且展现出良好的扩展性。

With the commercialization of Intel Optane DC persistent memory module(DCPMM)and the decrease in re-mote direct memory access(RDMA)hardware costs,the design of key-value(KV)stores that integrate non-vola-tile memory(NVM)and RDMA faces new opportunities and challenges.The key to building a KV store based on NVM and RDMA lies in designing an efficient communication protocol.Unfortunately,existing work either uses an NVM-oblivious RDMA protocol or an inefficient NVM-aware RDMA protocol,which prevents them from maximizing the performance of KV stores.This paper proposes the BOOM protocol,a new type of NVM-aware RDMA protocol.Compared to NVM-unaware protocols,the BOOM protocol allows direct RDMA operations on remote NVM,elimi-nating redundant data copies.Compared to existing NVM-aware protocols,it can significantly reduce metadata re-quests and reduce end-to-end latency of KV requests.Based on the BOOM protocol,BOOM-KV is built and further optimized for server central processing unit(CPU)utilization and crash persistence.Finally,BOOM-KV is com-pared with the state-of-the-art protocols,and results show that BOOM-KV can significantly reduce request latency with a maximum reduction of 42%in PUT latency and 41%in GET latency,exhibiting good scalability.