NAVIS boosts vector database performance by 2.74x
New on-SSD vector search system cuts insertion time by 2.74x while reducing latency by 25%
Deep Dive
A new on-SSD graph-based vector search system called NAVIS enhances average insertion throughput by up to 2.74x and average concurrent search throughput by up to 1.37x while reducing average search latency by up to 25.26%. It achieves this by driving down position-seeking overhead through a layout-supported selective vector read, a dynamic lightweight entrance graph update mechanism, and an entrance graph-aware edgelist cache.
Key Points
- NAVIS boosts insertion throughput by 2.74x and search throughput by 1.37x in on-SSD vector databases
- Uses selective vector reads, dynamic entrance graphs, and edgelist caching to cut position-seeking overhead
- Reduces search latency by 25.26% while handling high-dimensional workloads efficiently
Why It Matters
NAVIS accelerates vector database operations by 2.74x, enabling real-time AI applications to scale without performance degradation.