Secure and Efficient RDMA NIC Cryptography Offloading for Memory Disaggregation#
ACM CoNEXT 2025 digest Program Core pick 1–4 December 2025
Wonsup Yoon; Sue Moon (KAIST)
- Why: Sits squarely at the intersection of two of your core areas — RDMA fabrics and security. Memory disaggregation is moving from research to production, and almost nobody encrypts RDMA traffic because the performance cost is brutal. This is one of the first systems to make encrypted RDMA practical by offloading crypto to the RNIC itself.
- What: A secure memory-disaggregation system providing confidentiality and integrity for data in transit and at memory nodes, with crypto offloaded to the RNIC. Reported result: up to 9.63× lower P99.9 latency than a CPU-based implementation on a RocksDB workload (from abstract only).
- How: Three key datapath designs — dynamic MKey reconfiguration, configuration-data batching, and MKey pooling — to work around RNIC hardware limits and slow on-NIC encryption. The central finding is that the win comes from overlapping CPU compute with RNIC encryption via a yielding mechanism that lets the CPU do other work while the NIC encrypts.
- Where applicable: Memory-disaggregated clusters with RDMA NICs that support MKey manipulation and crypto offload. Notably, the paper is candid that RNIC crypto is relatively slow — the design leans on overlap rather than raw crypto throughput. Hardware-generation-dependent; not directly portable to older RNICs.