Adjust openai_vector benchmark for gpu indexing 9.3

[mayya-sharipova]

• Replaces the merge-wait polling with an explicit force-merge operation
• Removes parallel search and indexing
• Optimizes indexing performance by disabling merge throttling and refresh intervals

---

For GPU indexing pass the following params.json file to a rally run:

params.json

{
  "vector_index_type" : "hnsw",   
  "initial_indexing_bulk_indexing_clients" : 8,
  "mapping_type" : "vectors-only-with-docid",
  "initial_indexing_corpora" : "openai_initial_indexing_base64",  
  "parallel_indexing_corpora": "openai_parallel_indexing_base64"
}

Run with:

./rally race ...--runtime-jdk=25 --track-params=params.json --on-error=abort

---

Rally setup

vim ~/.rally/rally.ini

...
[system]
env.name = local
passenv = PATH,LD_LIBRARY_PATH

---

An example of running openai_vector track

params.json

{
  "vector_index_type" : "hnsw",   
  "initial_indexing_bulk_indexing_clients" : 8,
  "initial_indexing_corpora" : "openai_initial_indexing_base64",  
  "parallel_indexing_corpora": "openai_parallel_indexing_base64"
}

./rally race --track-path /home/ubuntu/elastic/rally-tracks/openai_vector \
--pipeline from-sources \
--revision current \
--car "16gheap" \
--user-tags "tag:openai_vector_gpu" \
--report-file /home/ubuntu/elastic/benchmarks/rally_results/openai_vector_gpu.md \
--runtime-jdk=25   \
--track-params=/home/ubuntu/elastic/benchmarks/rally/params.json \
--on-error=abort

To compare with CPU based index disable gpu indexing with --car-params="additional_cluster_settings:{'vectors.indexing.use_gpu':true}":

./rally race --track-path /home/ubuntu/elastic/rally-tracks/openai_vector \
      --pipeline from-sources \
      --revision current \
      --car "16gheap" \
      --user-tags "tag:openai_vector_cpu" \
      --report-file /home/ubuntu/elastic/benchmarks/rally_results/openai_vector_cpu.md \
      --runtime-jdk=25 \
      --track-params=/home/ubuntu/elastic/benchmarks/rally/params.json \
      --car-params='{"additional_cluster_settings": {"vectors.indexing.use_gpu": false}}' \
      --on-error=abort

---

Troubleshoot

• Check ~/.rally/benchmarks/races/ /rally-node-0/logs/server/rally-benchmark.log. If you see similar output in the elasticsearch server logs:
  [o.e.x.g.GPUSupport ] [elasticsearch-0] Found compatible GPU [NVIDIA L4] (id: [0])

it means your setup is successful, and Elasticsearch can use GPU. If you don’t see this line from GPUSupport, you should see an explanation why elasticsearch can’t use your GPU device.