[LUCENE-10391] Reuse data structures across HnswGraph invocations

• Type: Task
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 9.1
• Labels: vector-based-search
• Lucene Fields: New
• Assignee: Julie Tibshirani
• Reporter: Adrien Grand
• Created: 26/Jan/22 17:02
• Updated: 28/Aug/22 16:35
• Resolved: 02/Mar/22 01:30

Description

Creating HNSW graphs involves doing many repeated calls to HnswGraph#search. Profiles from nightly benchmarks suggest that allocating data-structures incurs both lots of heap allocations (http://people.apache.org/~mikemccand/lucenebench/2022.01.23.18.03.17.html#profiler_1kb_indexing_vectors_4_heap) and CPU usage (http://people.apache.org/~mikemccand/lucenebench/2022.01.23.18.03.17.html#profiler_1kb_indexing_vectors_4_cpu). It looks like reusing data structures across invocations would be a low-hanging fruit that could help save significant CPU?

Issue Links

• links to: GitHub Pull Request #641