Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

I/O Optimizations for Graph-Based Disk-Resident Approximate Nearest Neighbor Search: A Design Space Exploration

Liang Li, Shufeng Gong, Yanan Yang, Yiduo Wang, Jie Wu

TL;DR

This work presents an I/O-first framework for disk-based ANN that organizes techniques along three dimensions: memory layout, disk layout, and search algorithm, and introduces a page-level complexity model that explains how page locality and path length jointly determine page reads.

Abstract

Approximate nearest neighbor (ANN) search on SSD-backed indexes is increasingly I/O-bound (I/O accounts for 70--90\% of query latency). We present an I/O-first framework for disk-based ANN that organizes techniques along three dimensions: memory layout, disk layout, and search algorithm. We introduce a page-level complexity model that explains how page locality and path length jointly determine page reads, and we validate the model empirically. Using consistent implementations across four public datasets, we quantify both single-factor effects and cross-dimensional synergies. We find that (i) memory-resident navigation and dynamic width provide the strongest standalone gains; (ii) page shuffle and page search are weak alone but complementary together; and (iii) a principled composition, OctopusANN, substantially reduces I/O and achieves 4.1--37.9\% higher throughput than the state-of-the-art system Starling and 87.5--149.5\% higher throughput than DiskANN at matched Recall@10=90\%. Finally, we distill actionable guidelines for selecting storage-centric or hybrid designs across diverse concurrency levels and accuracy constraints, advocating systematic composition rather than isolated tweaks when pushing the performance frontier of disk-based ANN.

I/O Optimizations for Graph-Based Disk-Resident Approximate Nearest Neighbor Search: A Design Space Exploration

TL;DR

This work presents an I/O-first framework for disk-based ANN that organizes techniques along three dimensions: memory layout, disk layout, and search algorithm, and introduces a page-level complexity model that explains how page locality and path length jointly determine page reads.

Abstract

Approximate nearest neighbor (ANN) search on SSD-backed indexes is increasingly I/O-bound (I/O accounts for 70--90\% of query latency). We present an I/O-first framework for disk-based ANN that organizes techniques along three dimensions: memory layout, disk layout, and search algorithm. We introduce a page-level complexity model that explains how page locality and path length jointly determine page reads, and we validate the model empirically. Using consistent implementations across four public datasets, we quantify both single-factor effects and cross-dimensional synergies. We find that (i) memory-resident navigation and dynamic width provide the strongest standalone gains; (ii) page shuffle and page search are weak alone but complementary together; and (iii) a principled composition, OctopusANN, substantially reduces I/O and achieves 4.1--37.9\% higher throughput than the state-of-the-art system Starling and 87.5--149.5\% higher throughput than DiskANN at matched Recall@10=90\%. Finally, we distill actionable guidelines for selecting storage-centric or hybrid designs across diverse concurrency levels and accuracy constraints, advocating systematic composition rather than isolated tweaks when pushing the performance frontier of disk-based ANN.
Paper Structure (35 sections, 3 equations, 24 figures, 7 tables)

This paper contains 35 sections, 3 equations, 24 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Nearest Neighbor Search
  4. Disk-Based ANN
  5. I/O is the Main Bottleneck
  6. I/O Challenges Analysis
  7. I/O Complexity for Best-First Search
  8. Objectives
  9. Design Space
  10. Memory Layout Optimization
  11. Product Quantization (PQ)
  12. Cache Management (Cache)
  13. Hierarchical Graphs (MemGraph)
  14. Disk Layout Optimization
  15. Page Shuffle (PS)
  16. ...and 20 more sections

Figures (24)

  • Figure 1: DiskANN data layout. Each record is page-aligned and packed into 4 KB pages.
  • Figure 2: Latency breakdown across four datasets.
  • Figure 3: Quantization reduces disk reads by moving most distance calculation to memory.
  • Figure 4: Cache miss illustration.
  • Figure 5: Entry-point optimization. An in-memory navigation graph supplies high-quality entry points for disk-based search.
  • ...and 19 more figures