A Study Of Data Persistence Mechanism Based On Remote Direct Memory Access

In recent years,the emergence of Persistent Memory(PM)and Remote Direct Memory Access(RDMA)technologies have excited increasing interests in using them in datacenter environments,due to the persistence characteristics of PM and the low-latency performance of RDMA.However,the combination of PM and RDMA poses a significant challenge for distributed systems to guarantee remote data persistence because of the volatile cache in RDMA Network Interface Card(RNIC).Durable RDMA operations and the visibility of RDMA updates have not been fully explored in remote persistent memory systems.Although a few Remote Procedure Calls(RPC)-based research support remote data persistence via a set of RDMA operations,those software-based persistent solutions require remote CPU intervention and postpone the visibility of remote persistence for the sender,which leads to increased transmission latency.To avoid the shortcomings of existing RPC designs,proposed persistent RPCs use several hardware-supported RDMA Flush primitives to flush data from the volatile cache of RNIC to the persistent region for failure recovery.These persistent RPCs can decouple the data persisting from the complicated RPC processing through batch optimization and concurrent transmission of confirmed persistent RPC requests.In addition,persistent RPCs build persistent buffers at the receiver to store requests and RDMA connection information as redo logs to provide fast failure recovery.The multiple implementation of persistent RPCs can also meet the requirements of various system environments and application settings.A comparative study of RDMA-based RPC designs can clearly demonstrate the impacts of RPC designs on transmission performance.And,in a real testbed equipped with Intel Optane DCPMM and Infini Band network,the comparisons of persistent RPC designs based on RDMA Flush primitives with several state-of-the-art RPC designs will reflect the performance benefits of the persistent RPC design more authoritatively.Experimental results show that the RDMA Flush primitives and RPC designs can improve the throughput of RPCs by up to 90%,and reduce the 99 th percentile latency by up to 49%.The experimental studies also provide substantial and instructive guidelines for designing RDMA-based distributed PM systems.