2026 Volume E109.B Issue 4 Pages 634-647
Remote Direct Memory Access (RDMA) is a key technology in Data Center Networks (DCN) owing to its low latency, high bandwidth, and Central Processing Unit (CPU) bypass capabilities. However, RDMA over Converged Ethernet version 2 (RoCEv2) is suitable for standard RDMA transmission, making it difficult to utilize the abundant parallel resources in DCNs. While multipath transmission enhances bandwidth, it introduces out-of-order (OoO) packets. Recent studies have confirmed that even a small fraction of OoO packets can significantly degrade RDMA throughput due to the Go-Back-N (GBN) retransmission mechanism. OoO events frequently occur in practical DCNs when flowlets are split too aggressively or packets are sprayed across multiple paths, making OoO handling a critical problem. Moreover, existing bitmap-based solutions suffer from low memory efficiency and high processing delays. To address these problems, this study proposes TRNIC, an optimized bitmap design for fast and efficient packet reordering. TRNIC adopts a triple-table structure for bitmap sharing and memory efficiency, an array structure for efficient random access, and scheduling isolation for concurrent Work Queue Elements (WQEs) within a single Queue Pair (QP). The proposed TRNIC innovatively presents a high-performance OoO packet handling solution based on bitmap for the first time. Compared with Mellanox CX5 and Xilinx ERNIC, our proposed TRNIC significantly improves RDMA throughput under OoO conditions, achieving nearly 97 Gbps, and greatly reduces Flow Completion Time (FCT) compared to GBN.
