vectrust

Vectrust Benchmarks & Performance Analysis

This document provides benchmark results, methodology, tuning guidelines, and performance trade-offs for Vectrust.

---

1. Test Environment

• Operating System: Windows 11 (x86_64)
• Rust Toolchain: Rust 2024 Edition (cargo 1.99.1)
• Compilation Flags: --release (opt-level = 3, lto = "thin", codegen-units = 1)
• Dataset: Synthetic uniform vectors ($D = 64$ dimensions, Cosine Metric)

---

2. HNSW vs Flat Index (5,000 Vectors, Dim=64)

Running cargo run --release --example hnsw_vs_flat_benchmark:

Metric	Flat Index (Exact)	HNSW Index (ANN)	Improvement
Search Latency (Avg)	683.52 µs / query	297.71 µs / query	2.30x faster
Total Query Time (100 Qs)	68.35 ms	29.77 ms	2.30x faster
Recall@10 Accuracy	100.0% (Ground Truth)	93.00%	Sub-millisecond ANN
Batch Insertion Time	8.54 ms	1.96 s	Trade-off: Graph Build vs Query

Note: For large vector collections ($N > 100,000$), HNSW exhibits logarithmic $O(\log N)$ scaling, while Flat scales linearly $O(N)$. At 100,000 vectors, HNSW is typically 30x - 50x faster than Flat brute force.

---

3. High-Concurrency Search Throughput

Running cargo run --release --example concurrent_search:

• Workload: 4 concurrent search worker threads + 2 concurrent ingestion writer threads
• Dataset: 2,000 vectors baseline (Dim=32) + continuous streaming inserts
• Throughput: > 8,500 Queries Per Second (QPS) sustained on consumer hardware
• Zero Lock Poisoning: parking_lot::RwLock eliminates lock contention among read threads.

---

4. HNSW Parameter Tuning Guide

Parameter	Default	Range	Impact & Trade-offs
m	16	8 - 64	Max outgoing edges per node in upper layers. Higher values improve recall and graph connectivity on high-dimensional vectors, but increase memory footprint and build time.
m0	32	16 - 128	Max connections at base Layer 0. Typically set to $2 \times M$.
ef_construction	100	40 - 512	Exploration beam width during index construction. Increasing improves graph quality and recall, but increases insertion latency.
ef_search	64	16 - 256	Exploration beam width during query execution. Can be overridden per-query with query.with_ef_search(ef). Higher values increase recall at the cost of slight latency.

Tuning for Ultra-Low Latency (< 100 µs):

let config = HnswConfig {
    m: 12,
    m0: 24,
    ef_construction: 64,
    ef_search: 24,
};

Tuning for Maximum Recall (> 98%):

let config = HnswConfig {
    m: 24,
    m0: 48,
    ef_construction: 200,
    ef_search: 128,
};

This site is open source. Improve this page.