Highly Concurrent RDMA-based One-sided Access To NVM Mechanism Research

On one hand,in nowadays data center technologies based on network connection,one specific network access technique called RDMA(Remote Direct Memory Access)is widely adopted.RDMA can bypass the CPU-involving part of the traditional network access stack,directly accessing the memory of remote machines,and hence improves the network performance significantly.On the other hand,a new kind of main memory technology called NVM(Non-Volatile Memory)has drawn the attention from both industrial community and academic circles.NVM assures non-volatility(which means it can store data across power loss)while preserving traditional memory characteristics,including byte-addressability and relatively high access speed.Besides,NVM has cheaper price and high volumn density.Therefore,NVM could be utilized as memory device to replace the memory and hard drive in data centers partially or totally for much larger volumn and better performance,gaining great popularity in recent years.Recently,certain research aiming at further utilizing RDMA emerge.These research focus on how to utilize the one-sided characteristic of RDMA,re-constructing the traditional data structures and algorithms of network systems for one-sided-access-based network communication instead of message-passing-based network communication,and such novel network access paradigm could be called server-bypass paradigm.However,utilizing one-sided RDMA in such way brings the challenge of concurrent one-sided access,and compromises are made in previous researches.Also,researches begins to try combining both RDMA and NVM together,to gain benefits from both one-sided RDMA and NVM and retrieve even better performance.However,applying one-sided RDMA on NVM brings the challenge of NVM data crash consistency,due to the remote unawareness of one-sided RDMA and the non-volatility of NVM.Currently there are limited solutions on the challenge.In this paper,we propose a novel server-bypass RDMA to NVM framework,named Forca,to provide high concurrency and guarantee remote crash consistency at the same time.In Forca,we design an optimized log-structured mechanism to eliminate in-place updates,removing race conditions of one-sided RDMA and providing atomicity for each update simultaneously.We implement Forca as an generic module to support serverbypass RDMA to PM.For evaluation,we build a key-value store based on Forca,called ForcaKV,and conduct experiments on it.The experiment results show that Forca achieves higher throughput than state-of-the-art techniques by up to 1.4x under high concurrency scenarios,and its crash consistency mechanism incurs only 14.7%-19.1% time overhead.