Network Load Balancing With No Out-of-Order Packet for RoCE

Remote Direct Memory Access (RDMA) is becoming a popular high-speed network technology. It uses kernel bypass and zero copy to allow the host to exchange data directly in memory, providing high throughput and low latency while reducing CPU overhead. However, RDMA is based on single-connection transmission and does not make good use of the end-to-end path diversity of the data center. At the same time, existing load balancing algorithms cannot be well applied to RDMA. In this paper, we focus on the flow characteristics of RDMA and load balancing algorithm for RDMA. We first observe the flow characteristics of RDMA and the transmission mode in the lossless environment through experiments, and analyze the reasons for this phenomenon. Then we propose Ursine, a load balancing algorithm for RDMA. It achieves fine-grained RDMA load balancing while avoiding packet out-of-order, and effectively utilizes the end-to-end path diversity in the data center. We implemented the prototype based on DPDK and verified through small-scale experiments that Ursine provides better performance than existing algorithms. Finally, we analyze and look forward to the future and development trends of RDMA load balancing.