POSTER: CAVER: Enhancing RDMA Load Balancing by Hunting Less-Congested Paths

Remote Direct Memory Access (RDMA) has become a prevailing technology for modern data centers (DCs) to achieve high throughput and low latency 8. Many DCs have adopted RDMA over Converged Ethernet v2 (RoCEv2) 3 to provide superior performance for emerging application paradigms such as cloud storage 4 and distributed deep learning 11. Network load balancing (LB) plays a critical role in optimizing the DC network performance. There is a large body of literature studying LB for traditional DCs 1, 2, 6, 7, 9, as it is well-known that the widely-used ECMP 12 has a limited LB performance. However, RDMA operates in a different manner compared to traditional TCP-based data transmission, and existing studies for traditional DCs do not well fit RDMA-enabled DCs. For example, RDMA is very sensitive to out-of-order packets which may lead to significant throughput degradation, and also, RDMA flow can hardly be partitioned into flowlets. Thus, existing packet-level LB approaches 6, 7 and flowlet-level LB approaches 2 perform poorly in RDMA-enabled DCs.