btw this article[1] by Tom Burgmans is relevant. It is posted on linkedin though and never came into my feed. I found it on the relevance-search slack channel. [1] https://www.linkedin.com/pulse/testing-dense-vector-search-scale-part-1-ann-tom-burgmans-0tale/
On Thu, Mar 28, 2024 at 3:47 PM rajani m <rajinima...@gmail.com> wrote: > @Alessandro, > Is there a solr blog site where we can submit work/articles or are you > suggesting to post on my own site and share a link here? I prefer the > former if there is one because there were times when I had my own, > it hardly had any views and on top of that google blogging made me migrate > from blogs to sites and sites got deprecated. Is there or can we have a > solr specific wiki/blog site where solr users can submit common features > configs/modules configs/examples/performance metrics and so on....and maybe > have a voting/likes to confirm it works. We will have one common place to > submit and look for. > > > > On Thu, Mar 28, 2024 at 3:33 PM rajani m <rajinima...@gmail.com> wrote: > >> Run the same knn queries at a slow throughput for 30-60 minutes, this >> should warm up disk caches with hnsw index files, and then you should see a >> significant drop in the query time. Also make use of "fq" and reduce the >> document space as much as you can. >> >> On Thu, Mar 28, 2024 at 12:50 PM Iram Tariq >> <iram.ta...@northbaysolutions.net.invalid> wrote: >> >>> Hi Alessandro, >>> >>> Thank you for the feedback. Kindly see my comments below, >>> >>> *Ale*: >>> https://www.elastic.co/blog/accelerating-vector-search-simd-instructions, >>> I >>> suggest to experiment with simD vector improvements (unless you are >>> already doing it) >>> >>> * We will try this soon. * >>> >>> *Ale*: What about the machine memory? >>> >>> Following is the system specification: Linux ( CPU:64, RAM:488 GB, >>> OS:Ubuntu 20.04.6 ) >>> >>> *Ale*: you can fine-tune the hyper-parameter to compromise a bit on >>> recall >>> in favour of performance (hnswBeamWidth, hnswMaxConnections) >>> >>> I am trying this as a first step. But I am sure it will impact recall. >>> >>> Regards, >>> >>> >>> Iram Tariq | Software Architect >>> >>> NorthBay >>> >>> Direct: +1 (902) 329-7329 >>> >>> iram.ta...@northbaysolutions.net >>> >>> www.northbaysolutions.com >>> >>> >>> >>> >>> On Thu, Mar 28, 2024 at 5:42 AM Alessandro Benedetti < >>> a.benede...@sease.io> >>> wrote: >>> >>> > That's interesting. >>> > I think it's vital to get back some performance tests from the >>> community. >>> > Since my contribution to support Vector-search in Apache Solr was >>> merged, >>> > we got little or null feedback to understand its performance, in >>> real-world >>> > use cases. >>> > Blogs, open benchmarks or even just this sort of mail message are >>> welcome. >>> > Let me reply in line: >>> > -------------------------- >>> > *Alessandro Benedetti* >>> > Director @ Sease Ltd. >>> > *Apache Lucene/Solr Committer* >>> > *Apache Solr PMC Member* >>> > >>> > e-mail: a.benede...@sease.io >>> > >>> > >>> > *Sease* - Information Retrieval Applied >>> > Consulting | Training | Open Source >>> > >>> > Website: Sease.io <http://sease.io/> >>> > LinkedIn <https://linkedin.com/company/sease-ltd> | Twitter >>> > <https://twitter.com/seaseltd> | Youtube >>> > <https://www.youtube.com/channel/UCDx86ZKLYNpI3gzMercM7BQ> | Github >>> > <https://github.com/seaseltd> >>> > >>> > >>> > On Wed, 27 Mar 2024 at 21:06, Kent Fitch <kent.fi...@gmail.com> wrote: >>> > >>> > > Hi Iram, >>> > > >>> > > Is the machine doing lots of IO? If the hnsw graphs are not entirely >>> in >>> > > memory, performance will be poor. What JVM? You may get some benefit >>> from >>> > > simd support in java 21. Can you use the latest quantisation changes >>> in >>> > > Lucene to reduce memory footprint of the hnsw graphs? That's a large >>> > topk, >>> > > but I guess you need it? >>> > > >>> > > Best regards >>> > > >>> > > Kent Fitch >>> > > >>> > > On Thu, 28 Mar 2024, 5:12 am Iram Tariq, >>> > > <iram.ta...@northbaysolutions.net.invalid> wrote: >>> > > >>> > > > Hi All, >>> > > > >>> > > > I am using Dense vectors in SOLR and facing slowness in it. Each >>> search >>> > > is >>> > > > taking 10-25 seconds. I want to reduce the time to 5 seconds (or >>> less >>> > > > ideally). >>> > > > >>> > > > Following configurations are being used. >>> > > > >>> > > > >>> > > > 1. *SOLR Version:* 9.3.0 >>> > > > 2. *Lucene Version:* 9.7.0 >>> > > >>> > *Ale*: >>> > >>> https://www.elastic.co/blog/accelerating-vector-search-simd-instructions >>> , >>> > I >>> > suggest to experiment with simD vector improvements (unless you are >>> > already doing it) >>> > >>> > > > 3. *Vector Dimensions*: 384 >>> > > > 4. *Total Shards:* 5 >>> > > > 5. *Number of Vectors (Per shard*): 43209158 >>> > > > 6. *JVM for each Instance:* 35GB >>> > > >>> > *Ale*: What about the machine memory? >>> > >>> > > > 7. *TopK: *1000 (Getting 1000 from each shard) >>> > > > 8. *Rows: *1000 >>> > > > 9. *Vector Field Schema: *<fieldType name="knn_vector_384" >>> > > > class="solr.DenseVectorField" hnswMaxConnections="20" >>> > > > knnAlgorithm="hnsw" >>> > > > vectorDimension="384" similarityFunction="cosine" >>> > hnswBeamWidth="40"/> >>> > > >>> > *Ale*: you can fine-tune the hyper-parameter to compromise a bit on >>> recall >>> > in favour of performance (hnswBeamWidth, hnswMaxConnections) >>> > >>> > > > 10. *Stored*: False >>> > > > 11. *WebServer:* Apache Tomcat >>> > > > 12. *System Specs*: Linux ( CPU:64, RAM:488 GB, OS:Ubuntu >>> 20.04.6 ) >>> > > > >>> > > > Any sort of help/clue will be appreciated. >>> > > > >>> > > > >>> > > > >>> > > > Regards, >>> > > > >>> > > > >>> > > > Iram Tariq | Software Architect >>> > > > >>> > > > NorthBay >>> > > > >>> > > > Direct: +1 (902) 329-7329 >>> > > > >>> > > > iram.ta...@northbaysolutions.net >>> > > > >>> > > > www.northbaysolutions.com >>> > > > >>> > > >>> > >>> >>
