DALdex: A DPU-Accelerated Persistent Learned Index via Incremental Learning
Aoyang Tong, Yu Hua, Menglei Chen · 2025
The Data Process Unit (DPU) specializes in offloading CPUintensive tasks and provides efficient fault tolerance through hardware-level isolation.This brings unique opportunities to develop persistent indexes with high performance and availability in High Performance Computing (HPC) systems.The recent learned index exploits machine learning models to efficiently fit data distributions, exhibiting superior performance and low storage costs, which is a promising alternative to traditional tree-based range indexes.However, state-of-the-art persistent learned indexes suffer from costly model retrainings and inefficient recovery mechanisms based on Non-Volatile Memory (NVM), making them inefficient to be offloaded to DPUs.To address these challenges, we propose DALdex, a CPU-DPU hybrid persistent learned index with high performance and availability.To mitigate model retraining overheads, DALdex offloads retraining tasks to DPU based on the incremental learning scheme.To minimize NVM amplifications, DALdex designs an NVM-friendly index structure that is decoupled into a DRAM-accelerated model layer and an NVM-aware block layer.Moreover, DALdex utilizes the hardware isolation feature of DPU to achieve seamless failover and instant recovery via the PCIe bus.Extensive evaluation results demonstrate that DALdex outperforms state-of-theart persistent indexes by 1.07-6.34×with minimal DRAM and NVM overheads.The open-source code of DALdex is available at https://github.com/CitySkylines/DALdex.