[CUBVEC-147] Disk-based HNSW Performance: Co-locate Node and Vector Data

[hgryoo]

http://jira.cubrid.org/browse/CUBVEC-147

Purpose

In CUBVEC-138, the disk-based HNSW vector index was initially implemented by storing
graph node metadata (neighbor links) and vector data in separate pages.

While this design simplified the initial implementation, it introduced noticeable performance overhead during search due to:

• Additional I/O required to access vector pages after loading graph nodes
• Increased page buffer fix/unfix operations
• Poor cache locality during graph traversal

Implementation

This PR modifies the storage layout to store vector data together with HNSW node metadata in the same page (or block).

Changes:

• struct node_t: Co-locate graph node information and its corresponding vector payload
• get_vector_by_slot_id ()
  • Eliminate extra page accesses for vector retrieval during search
  • Reduce page buffer operations and improve cache locality

Remarks

N/A