Accelerating Modern Database Systems by Leveraging Computational Storage and Cache-Coherent Interconnects
Sajjad Tamimi, Arthur Bernhardt, Ilia Petrov, Andreas Koch · IEEE Micro · 2026
Near-Data Processing (NDP) offers a promising approach to improve the performance of modern databases with hybrid workloads comprising long-running analytical and frequent short update transactions. This paper introduces an end-to-end NDP-capable system that integrates a computational storage device (CSD) with a database capable of offloading queries to storage while handling concurrent queries on the host. A novel host-device protocol combines high-throughput PCIe-DMA with low-latency messaging via cache-coherent interconnects (CCI) such as CXL or CCIX. The paper presents two systems, DANSEN and CINDA, each featuring a custom embedded NDP engine. DANSEN offloads read-only queries while ensuring consistency with concurrent host updates. CINDA extends this capability to update queries through a low-latency synchronization mechanism and a CCI-enabled communication protocol.