Gemma: Robust and Path-aware Loading Balancing in RDMA Networks

Load balancing in Remote Direct Memory Access (RDMA) networks is critical to network performance. Despite numerous load balancing algorithms devised to optimize multiple end-to-end links in data centers, they falter in RDMA networks due to unawareness of out-of-order packets and PFC pausing. Motivated by the above two issues, this paper proposes Gemma, a robust and path-aware load balancing algorithm. Gemma prevents out-of-order packets by proactively suspending and resuming the transmission of switch queues, thereby avoiding the performance degradation caused by retransmissions of out-of-order packets. In addition, Gemma leverages egress queue length as a path state indicator, integrating synchronized lengths from local and remote switches to comprehensively assess path congestion and facilitate prompt, effective rerouting to prevent frequently triggering PFC pausing. Moreover, Gemma can be extended to the Fat-Tree topology with a module that facilitates per-hop rerouting decisions on uplink paths, thereby enhancing scalability. Experiments show that compared to ConWeave, CONGA, and ECMP, Gemma achieves a 6%, 35%, and 57% improvement for average FCT under 80% load, which fully demonstrates the efficiency and feasibility of our proposed scheme.