Designing hash tables for RDMA: Challenges and optimization paths

Hash tables are fundamental for databases, key-value stores, and network systems, but as IoT data volumes explode, single-server memory becomes insufficient. Remote Direct Memory Access (RDMA) allows one machine to directly read/write another's memory at low latency, making disaggregated memory systems attractive. However, deploying hash tables on RDMA introduces three major challenges, as outlined in a new survey paper by Shuchen She and Haipeng Dai, accepted at the 20th China Conference on Wireless Sensor Networks (CWSN 2026).

First, a single logical hash table access may require multiple RDMA requests due to collision handling and probing, quickly exhausting NIC queue resources. Second, because RDMA bypasses the remote CPU, traditional concurrency control (e.g., locks, read-copy-update) that relies on remote threads no longer works, forcing designers to rethink synchronization. Third, RDMA NICs have limited memory registration, cache, and atomic operation capabilities, constraining hash table structures. The paper categorizes existing RDMA-friendly hash table designs and suggests promising directions such as using batching to reduce request counts, leveraging hardware atomic operations for concurrency, and designing locality-aware structures to minimize cache misses. For IoT big-data scenarios, these optimizations could enable scalable, low-latency in-memory storage across clusters.