[GitHub] [lucene] jtibshirani edited a comment on pull request #312: LUCENE-10109: Bump default beam width for HNSW

[GitBox <gi...@apache.org>]

jtibshirani edited a comment on pull request #312:
URL: https://github.com/apache/lucene/pull/312#issuecomment-923615530


   Here are some example results on the sift-128-euclidean dataset from ann-benchmarks. The recall/ QPS curve doesn't improve a huge amount after beamWidth passes 100.
   
   **M=16, beamWidth=16**
   Built index in 172.130s
   
   ```
   Approach                   Recall    QPS
   LuceneHnsw(n_cands=50)     0.672     4727.125
   LuceneHnsw(n_cands=100)    0.774     3071.469
   LuceneHnsw(n_cands=500)    0.927      807.171
   ```
   
   **M=16, beamWidth=100**
   Built index in 881.74s
   ```
   Approach                   Recall    QPS
   LuceneHnsw(n_cands=50)     0.878     3876.651
   LuceneHnsw(n_cands=100)    0.941     2347.171
   LuceneHnsw(n_cands=500)    0.993      612.210
   ```
   
   **M=16, beamWidth=500**
   Built index in 3156.50s
   ```
   Approach                   Recall    QPS
   LuceneHnsw(n_cands=50)     0.891     3737.885
   LuceneHnsw(n_cands=100)    0.952     2229.307
   LuceneHnsw(n_cands=500)    0.996      571.994
   ```
   
   The [HNSW paper](https://arxiv.org/abs/1603.09320) also cites 100 as a reasonable example for efConst: "... for a 10M SIFT dataset and shows that a reasonable quality index can be constructed for efConstruction=100 on a 4X 2.4 GHz 10-core Xeon E5-4650 v2 CPU server in just 3 minutes. Further increase of the efConstruction leads to little extra performance but in exchange of significantly longer construction time."


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org



---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org