LBCB: One-sided RDMA-based Distributed B+ tree Index with Low Bandwidth Consumption
Jibo Liu, Rui Xi, Qinzhen Cao, Xiaowen Nie, Zhuohan Hou · 2024
Disaggregated memory architecture segregates computing and memory resources into distinct pools interconnected by a high-speed one-sided RDMA (Remote Direct Memory Access) network, enhancing memory utilization, reducing costs, and facilitating elastic scaling of computing and memory resources. However, optimizing index structures to maximize the benefits of this framework poses significant challenges. Despite progress in B+ tree indexes for disaggregated memory systems, they still suffer from severe write and read amplification issues, which impede latency and throughput optimizations.In this paper, we propose LBCB, a B+ tree index for disaggregated memory, which significantly reduces the bandwidth consumption of index operations. First, LBCB introduces an RDMA friendly B+ tree leaf node structure, which improves concurrency while reducing the bandwidth consumption. Second, LBCB designs a logical fusion FAA lock that synchronizes more information with a single RDMA communication, significantly reducing the number of RDMA network round trips. Finally, LBCB optimistically compresses the critical path of index operations, further optimizing latency and throughput. Evaluation results show that compared with the state-of-the-art distributed B+ tree index, the write performance of LBCB is improved by 1.42 times.