milvus

This work addresses the challenges faced by existing vector retrieval systems in large-scale real-time scenarios, where low throughput, difficulty in dynamic updates, and GPU memory constraints hinder the simultaneous achievement of high accuracy and low latency. To overcome these limitations, we propose a CPU-GPU-disk协同 real-time vector retrieval framework that integrates a hierarchical indexing structure, workload-aware caching, CUDA multi-stream optimization, concurrency control, and adaptive resource scheduling to enable efficient online updates and low-latency queries. Experimental results across diverse streaming workloads demonstrate that our system achieves an average 20.9× throughput improvement and reduces query latency by 1.3–50.7× while maintaining high recall, thereby significantly breaking through the performance bottlenecks of dynamic large-scale vector retrieval.

This work addresses the performance bottleneck in existing relational databases that execute vector search on CPUs, thereby failing to leverage the parallel computing capabilities of GPUs. To bridge this gap, the authors extend the TPC-H benchmark with vector data and develop a modular query engine supporting hybrid CPU/GPU execution. They propose GPU-optimized vector indexing and embedding organization strategies, complemented by heterogeneous memory management and evaluation across multiple index structures and interconnect architectures (e.g., PCIe and NVLink). The study reveals that relational operations benefit more significantly from GPU acceleration than vector search alone. Under high-speed interconnects, the proposed optimizations enable synergistic acceleration of both relational and vector components on the GPU, substantially outperforming current database architectures.

To address the challenges of large-scale vector graphs exceeding single-GPU memory capacity—causing frequent GPU stalls, low retrieval throughput, and high latency in conventional CPU-GPU architectures—this paper proposes a pipelined approximate nearest neighbor (ANN) search system for GPU clusters. Our method introduces an asynchronous data transfer mechanism leveraging GPUDirect Async to overlap computation with communication, and establishes a distributed vector index management framework coupled with pipelined parallel computation to support ultra-large-batch queries. Experimental results demonstrate that, while maintaining high recall (>95%), our system achieves a 3.2× throughput improvement and reduces end-to-end latency by 67%, significantly outperforming state-of-the-art CPU-GPU and single-GPU baselines.

This work addresses the challenge of achieving efficient GPU-accelerated approximate nearest neighbor search (ANNS) by simultaneously optimizing index construction speed, query throughput, recall accuracy, and memory footprint. The authors propose an end-to-end GPU-native ANNS pipeline that integrates IVF clustering with RaBitQ—a scalable low-bit quantization method—and introduces a fused search kernel that eliminates the need to access original vectors during reranking. This design significantly improves the trade-off between efficiency and accuracy. Integrated into the NVIDIA cuVS library, the proposed approach achieves 2.2× higher queries per second (QPS) than CAGRA at a recall of approximately 0.95, while constructing the index 7.7× faster. Compared to IVF-PQ, it delivers over 2.7× higher throughput without requiring reranking on original vectors.

To address the prohibitively long indexing time and poor CPU-architecture compatibility of graph-based indexes in high-dimensional approximate nearest neighbor search (ANNS), this paper proposes Flash—a hardware-aware compact encoding strategy. Flash innovatively integrates vector quantization with key CPU architectural features, including SIMD parallelism and cache locality, enabling efficient distance computation while maintaining bounded quantization error. By jointly optimizing compact encoding, memory access patterns, and cache-friendly graph construction, Flash achieves 10.4×–22.9× speedup in index construction across eight real-world datasets ranging from 10 million to one billion vectors. Crucially, this acceleration comes without sacrificing retrieval accuracy or query latency—indeed, both are preserved or improved. Flash thus bridges the gap between algorithmic efficiency and modern hardware utilization in large-scale ANNS.

This work addresses the bandwidth bottleneck in large-scale approximate nearest neighbor search (ANNS) under high-recall regimes, where full-precision vector re-ranking is essential but constrained by GPU memory capacity, often forcing data to reside in CPU memory or SSDs. To overcome this limitation, the authors propose the first integration of high-bandwidth 3D NAND flash (HBF) directly into the GPU package, enabling on-package residency of billion-scale full-precision vector databases. By co-designing a near-storage search unit with an optimized IVF-PQ pipeline, the architecture eliminates data migration overhead during re-ranking, effectively breaking the traditional memory wall. Compared to conventional GPU-DRAM and GPU-SSD systems, the proposed design achieves up to 20× higher throughput and up to 40× lower latency, simultaneously delivering high recall and high throughput.

This work addresses the high latency and service disruption caused by frequent index rebuilds in existing approximate nearest neighbor (ANN) methods under dynamic vector database updates. To overcome these limitations, the authors propose ACRONYM—a co-designed algorithm-hardware platform that leverages a data-distribution-agnostic XOR-and-Accumulate (XAC) systolic array encoder and Hamming-distance-based search, integrated with content-addressable memory (CAM) to enable in-memory parallel computation. A two-stage coarse-to-fine retrieval architecture circumvents CAM dimensionality constraints, allowing continuous, interruption-free updates. Evaluated on million-scale dynamic datasets, ACRONYM achieves over 90% recall, 8 million queries per second throughput, only 32 MB memory footprint, and 2.56 μJ per query energy efficiency—outperforming CPU-based HNSW by 400× and GPU-based FAISS-IVF by 80× in speed.

This study addresses the lack of systematic evaluation of hybrid search mechanisms that combine semantic retrieval with metadata filtering in existing vector databases. We propose a novel relevance metric, Global-Local Selectivity (GLS), construct MoReVec—the first benchmark dataset supporting filtered retrieval—and extend ANN-Benchmarks to enable unified evaluation of hybrid search performance. Through comprehensive experiments integrating diverse filtering strategies into FAISS, Milvus, and pgvector with IVFFlat and HNSW indexes, we demonstrate that engine-level algorithmic integration critically governs performance: Milvus achieves more stable recall via hybrid execution, pgvector’s optimizer often selects suboptimal query plans, and IVFFlat outperforms HNSW under low-selectivity queries. Our findings culminate in practical configuration guidelines that offer both theoretical insights and actionable recommendations for efficient hybrid search deployment.

Existing graph-based indexes for vector search overlook the spatial and temporal locality inherent in query streams, leading to redundant traversals and suboptimal efficiency. This work proposes CatapultDB, the first approach to dynamically model query locality within ANN graph indexes and inject lightweight “catapult” shortcut edges that connect high-frequency query regions to relevant target nodes. By doing so, CatapultDB selects superior starting points for queries without altering the underlying graph structure or search algorithm. The method is transparently compatible with filtered search, dynamic insertions, and disk-resident indexes. Experimental results on four skewed workloads show that CatapultDB achieves up to 2.51× higher throughput than DiskANN while matching or exceeding its recall, delivering efficiency comparable to LSH—without requiring index reconstruction or sacrificing functionality—and gracefully adapts to shifting query patterns.

This work addresses the memory bandwidth bottleneck in high-dimensional approximate nearest neighbor search (ANNS) on CPUs and GPUs, where conventional early termination mechanisms struggle to accelerate computation due to slow distance convergence. The authors propose a hardware-software co-design that integrates DIMM-level near-data processing (NDP) with a PCA-statistics-based feature-level early stopping mechanism, employing an estimate-and-correct strategy to accurately approximate full-dimensional distances. Additionally, they introduce bit-level dynamic floating-point compression and data-aware neighbor list mapping to substantially reduce memory access and communication overhead. Evaluated under strict accuracy constraints, the proposed system achieves 8.4× and 1.4× speedups over state-of-the-art CPU and GPU baselines, respectively, and outperforms the latest NDP accelerator, ANSMET, by 1.69×.