| RPC is a component widely used in distributed system, the requirement for its bandwidth, throughput and latency becomes higher than ever. The performance of traditional RPC components based on TCP/UDP socket has been facing more and more challenges. At present high speed InfiniBand network has been widely used. InfiniBand network supports RDMA, and RPC components can achieve better performance when they use RDMA technology which can efficiently support many distributed systems.RPC over RDMA in user-space is a component based on IB verbs interface. The zero copy and the kernel bypass technique provided by the IB verbs interface in user-space can provide high bandwidth and low latency of network transmission. Currently RPC components based on IB verbs use RDMA_READ to access remote memory. RDMA_WRITE can simplify the process of network interaction in RPC components. This optimization method is described in detail. At the same time, a scheme of memory management in RPC layer is proposed for the problem of memory registration cost. The main idea is to repeatedly use the registered memory block, thus reducing the times of memory registration. In many cases, the allocation and recovery of memory blocks involved in the transmission is dominated by upper level programs. It is necessary to extend the standard interface of ONC RPC. Memory management interface which allows the upper level program to use the memory blocks managed by the RPC layer is added.By comparing the test data before and after optimization and TI-RPC under the same conditions, the optimization results of the above two aspects are verified respectively. The results show that the optimization of the transmission model significantly improves the performance when servers return data to the clients. The optimization for memory registration makes RDMA over RPC component have better performance in the transmission of small blocks, and broaden its scope of application for the size of the transmission block. |
PDF Full Text Request