VeloANN

SSD-resident graph ANN for high-throughput vector search with locality-aware storage and asynchronous execution.

VeloANN is an SSD-resident graph indexing system for high-throughput approximate nearest-neighbor search.

Problem

Graph-based ANN indexes deliver strong search quality, but billion-scale vector datasets often exceed memory capacity. Moving the graph index to SSD reduces memory pressure, but naive SSD-resident traversal suffers from random I/O, page over-fetch, buffer-pool thrashing, and poor overlap between I/O and compute.

Core idea

VeloANN treats storage layout, buffering, and execution as one co-designed system. It uses affinity-aware layout to reduce unnecessary page reads, a record-level buffer pool to avoid hot/cold page interference, and a coroutine-based asynchronous runtime to overlap SSD I/O with graph traversal computation.

My role

Research contribution to the vector-search storage substrate line, focusing on the system-level design space for SSD-resident graph indexing and high-throughput retrieval.

Evidence

  • ArXiv paper
  • SSD-resident graph ANN design with locality-aware layout, record-level buffering, and asynchronous execution.

Impact

VeloANN reaches 0.92× in-memory throughput with 10% memory, and reports 5.8× higher throughput than disk-based prior systems, showing that SSD-resident ANN can narrow the gap to in-memory graph search while sharply reducing memory demand.