Transmission Optimization And Fault Tolerance Of Key-Value Store Based On RDMA

With the in-memory computing gradually becoming the mainstream big data analysis method,the memory key-value storage system is widely used due to its fast access speed and strong scalability.However,in the face of increasing demands for large-scale data processing,network performance has gradually become a major bottleneck for key-value storage systems.On the other hand,Remote Direct Memory Access(RDMA)technology allows the machine to read and write memory from the pre-registered storage area of another machine without involving remote CPU interrupts,enabling low latency,high throughput,and low CPU overhead provides an opportunity to eliminate the above bottlenecks.However,most existing RDMA-based key-value storage systems simply and directly apply RDMA,and lack the consideration of the key-value store system's own characteristics,making it difficult to fully exploit the transmission performance advantages of RDMA.In addition,in view of the common storage node failure problems of key-value storage systems,existing systems mainly adopt a multi-copy backup strategy or an erasure code fault-tolerance strategy.Affect the overall performance of the system.In response to the above problems of the memory key-value store,this paper proposes a transmission optimization scheme based on RDMA verb communication semantics,and establishes a fault-tolerant mechanism based on erasure codes and RDMA calculation offload,thereby improving transmission efficiency,reducing CPU overhead and fault-tolerant redundancy overhead.The specific contributions are as follows:(1)In view of the problem that RDMA transmission performance cannot be fully tapped,this paper uses a lower latency unilateral RDMA write semantics to redesign the GET instruction and PUT instruction of the memory key-value storage system,so that a round-trip transmission can complete data access GET instruction,which greatly reduces the transmission delay and releases the client CPU overhead.Aiming at the problem of remote read and write competition caused by the unilateral write operation bypassing the remote core,a monitoring thread and a priority queue are introduced in the server-side RDMA memory to convert multiclient parallel instructions into a polled serial processing scheme.(2)Aiming at the normalization of node failures in key-value storage systems,a faulttolerant mechanism based on erasure codes and RDMA calculation offload is established.The mechanism uses an erasure code scheme with less storage redundancy for data redundancy and verification,and for the problem of high CPU calculation brought by the use of erasure codes,the programmability of RDMA network card hardware is used to erasure Code encoding and decoding calculations are offloaded from the CPU to the network card,and work with the sender's write high-performance transmission to form a pipeline to work alternately,thereby liberating the sender's CPU resources and reducing the cost of server-side faulttolerant storage and recovery decoding.Experimental results based on real environments show that the proposed design can reduce latency by 32%,increase client throughput by 16%,reduce client CPU load by 60%,and reduce storage overhead by 65% compared to the original key-value store.