RDMA Acceleration Scheme for Long-Distance Optical Network

To accommodate large-scale data processing requiring high bandwidth, low latency, and low power consumption, remote direct memory access (RDMA) is extensively utilized in today’s data centers (DCs). However, due to the limitations in constructing and operating large centralized DCs, major DC operators are gradually transitioning towards distributed DCs comprised of multiple smaller DCs. As a result of that transition, RDMA-based workload processing across widely distributed DCs will be required in the future. The most commonly used transport mode in RDMA, reliable connection (RC) mode, adopts an ACK scheme to prevent packet loss; however, that scheme exacerbates the "long fat pipe" problem as transmission distance increases, so throughput is degraded. In consideration of the above-described situation, an RDMA WAN accelerator—which improves throughput by sending pseudo-ACKs to the sender node and reduces the ACK waiting time on the sender side— is proposed. Additionally, to restore the reliability of packet delivery compromised by pseudo-ACKs, a mechanism called "segment-adaptive retransmission control" is also implemented in the accelerator. The performance (i.e., throughput relative to transmission distance) of the proposed accelerator was evaluated by using actual servers and RNICs and compared with that of standard RDMA. In the case of data transfers of 4096byte messages over a network spanning 100 km, the proposed RDMA WAN accelerator improves performance by about 20 times compared to that of standard RDMA.