An Out-of-Order Packet Processing Algorithm of RoCE Based on Improved SACK

RoCE (RDMA over commodity Ethernet) combines the go-back-N retransmission mechanism, priority-based flow control (PFC) and congestion control (CC) algorithms to achieve the low latency, low CPU overhead and high bandwidth required by data center networks (DCN). However, when faced with scenarios such as multi-path and adaptive routing that may lead to out-of-order packets, RoCE is prone to serious throughput degradation. This paper proposes an improved Selective Acknowledgement (SACK) algorithm for RoCE (SACK-RoCE) to solve the above problem. The SACK-RoCE includes three mechanisms: packet trace, out-of-order packet detection, and lost packets retransmission. In simulation, we test the performance of the SACK-RoCE in both single flow and real-world flows environment. The SACK-RoCE improves the throughput by 7 times compared to RoCE and can almost keep the number of retransmission times consistent with the number of lost packets. As for real loads, the SACK-RoCE outperforms the improved RoCE NIC (IRN) by 12.05%∼22.32% on average FCT, tail FCT and average slowdown metrics. In addition, the SACK-RoCE adds only 315 bits of memory usage per link, which makes the algorithm easy to deploy.