PageANN

Overview

PageANN is a disk-based approximate nearest neighbor search (ANNS) framework designed for high performance and scalability. The framework enables efficient large-scale ANN search under various memory budgets while maintaining high recall accuracy.

Key Innovation

PageANN introduces a page-node graph structure that aligns logical graph nodes with physical SSD pages, where similar vectors are clustered into page nodes. This mapping ensures that each visited graph node corresponds to a single SSD page read.

Technical Approach

• Co-designed disk data layout leverages page-node structure
• Merging technique stores only representative vectors and topology information
• Avoids unnecessary reads through intelligent data organization
• Memory management strategy combines lightweight indexing with coordinated memory-disk allocation

Performance Results

Experimental results demonstrate significant advantages:

• 1.85x-10.83x higher throughput vs. state-of-the-art disk-based ANNS methods
• 51.7%-91.9% lower latency across different datasets and memory budgets
• Maintains comparable high recall accuracy
• Reduces random I/O through page-level graph organization
• Improves SSD utilization during vector search

Research

Published on arXiv (2509.25487) in late 2025. PageANN extends Microsoft's DiskANN with page-level graph organization.

Availability

Source code available on GitHub.