The rapid growth of Large Language Models (LLMs) has generated massive amounts of high-dimensional feature vectors extracted from diverse datasets. Efficient storage and retrieval of such data are critical for enabling accurate and fast query responses. Vector databases (Vector DBs) provide efficient storage and retrieval for high-dimensional vectors. These systems rely on Approximate Nearest Neighbor Search (ANNS) indexes, such as HNSW, to handle large-scale data efficiently. However, the original HNSW is implemented on DRAM, which is both costly and vulnerable to crashes. Therefore, we propose P-HNSW, a crash-consistent HNSW on persistent memory. To guarantee crash consistency, P-HNSW introduces two logs, NLog and NlistLog. We describe the logging process during the operation and the recovery process in the event of system crashes. Our experimental results demonstrate that the overhead of the proposed logging mechanism is negligible, while P-HNSW achieves superior performance compared with SSD-based recovery mechanisms.
The Industrial Internet of Things (IIoT) increasingly relies on vector embeddings for analytics and AI-driven applications such as anomaly detection, predictive maintenance, and sensor fusion. Efficient approximate nearest neighbor search (ANNS) is essential for these workloads. Graph-based methods are among the most representative methods for ANNS. However, most existing graph-based methods, such as Hierarchical Navigable Small World (HNSW), are designed for CPU execution on high-end servers and give little consideration to the unique characteristics of edge devices. In this work, we present ON-NSW, a GPU-optimized design of HNSW optimized for edge devices. ON-NSW employs a flat graph structure derived from HNSW to fully exploit GPU parallelism. In addition, it carefully places HNSW components in the unified memory architecture of NVIDIA Jetson Orin Nano. Also, ON-NSW introduces warp-level parallel neighbor exploration and lightweight synchronization to reduce search latency. Our experimental results on real-world high-dimensional datasets show that ON-NSW achieves up to 1.44× higher throughput than the original HNSW on the NVIDIA Jetson device while maintaining comparable recall. These results demonstrate that ON-NSW provides an effective design for enabling efficient and high-throughput vector search on embedded edge platforms.
With the rapid advancement of edge device performance, efficient on-device data management and processing have become critical. Transmitting data to edge servers introduces network overhead and a delayed data problem. However, modern edge devices now offer enhanced computing power and increased memory/storage capacity, enabling local data processing to address these challenges. In this work, we propose LION, a learned index for ondevice sensor data management. LION employs the OptimalPLR algorithm to train a linear model for each node. LION reduces training overhead by deferring the computation of the slope and intercept for each data node. Furthermore, to overcome limitations imposed by the maximum error bound, LION leverages Scope, a simple indirection table that improves lookup efficiency. Experimental results demonstrate that LION outperforms other indexing methods on both synthetic and real-world workloads.
Various devices such as smart cars, healthcare systems, and IoT platforms generate a massive amount of time series data. This type of data exhibits unique characteristics that pose new challenges for in-memory indexes, including heavy insert rates, monotonically increasing keys, and workloads dominated by range queries. In this paper, we propose a new in-memory index, Segmented Block-based Tree (SB-Tree), designed specifically for time series workloads. To support high insert throughput, SB-Tree decouples the index into a search layer and a data layer, and updates the search layer asynchronously to reduce write amplification. It introduces a shortcut mechanism to reduce tree traversal overhead and uses segmented blocks with per-thread data blocks to minimize contention during inserts. In addition, SB-Tree employs a lightweight block allocator to reduce system call overhead during memory management. We evaluate SB-Tree using synthetic time series workloads with varying levels of delayed data, inserting up to 1 billion keys across 80 threads. Our results show that SB-Tree outperforms state-of-the-art in-memory indexes, achieving up to $7\times $ higher throughput on insert-only workloads and $2.4\times $ lower 99.99th percentile tail latency on scan workloads compared to state-of-the-art.