Alibaba Cloud Elasticsearch AISearch performance whitepaper

Client ECS specifications

Server-side Elasticsearch specifications

Before you run the test, disable access logs to prevent logging overhead from affecting the results:

PUT /_cluster/settings
{
  "persistent": {
    "apack.accesslog.enabled": false,
    "apack.accesslog.search.enabled": false,
    "logger.index.search.accesslog": "OFF"
  }
}

Test datasets

The tests use Cohere10M as the primary dataset:

Set up the dual 16 vCPU test cluster

The scaling test runs on two 16 vCPU data nodes. Each data node holds a complete copy of the Cohere10M graph. Create the instance and configure the ES YML settings in the console before you run the stress test.

index.native.native_hnsw.search_worker_thread_count: "12"
index.native.native_hnsw.native_search_pool.enabled: "true"
index.native.native_hnsw.native_search_queue_capacity: "208"
index.native.native_hnsw.cpu_lending.mode: "static"

native.indices.queries.cache.enabled: "false"
native.indices.queries.cache.knn.enabled: "false"
indices.queries.cache.size: "0"

Create the native_hnsw index

Run the following request to create the cohere10m_native_hnsw test index:

PUT /cohere10m_native_hnsw
{
  "settings": {
    "index.havenask.engine.enabled": true,
    "index.sort.field": "id",
    "index.native.index.queries.cache.enabled": false,
    "index.native.index.queries.cache.knn.enabled": false,
    "index.requests.cache.enable": false
  },
  "mappings": {
    "_source": { "enabled": false },
    "properties": {
      "id": { "type": "long" },
      "vector": {
        "type": "dense_vector",
        "dims": 768,
        "index_options": {
          "type": "native_hnsw",
          "builder": {
            "add_chunk_size": 100000,
            "prune_headroom": 0.0
          }
        }
      }
    }
  }
}

The test index uses these native_hnsw graph parameters and contains one primary shard, one replica, and 10,000,000 documents. The primary shard and the replica reside on the two data nodes, and each shard has only one segment.

Query parameters

The Cohere10M tests use the following query parameters:

Run the Cohere10M test from the command line

The following commands run the two-level concurrency throughput test for Top10. Set the connection parameters, username, and password based on the actual configuration of your instance.

set -euo pipefail

export ES_SCHEME='http'
export ES_HOST='<Elasticsearch endpoint>'
export ES_PORT='9200'
export ES_INDEX='cohere10m_native_hnsw'
export ES_USER='<Elasticsearch username>'
export VDBBENCH_ES_QUERY_WIRE_FORMAT='cbor-f32le'
read -rsp 'Elasticsearch password: ' ES_PASSWORD
echo

VDB_COMMON=(
  --scheme "$ES_SCHEME" --host "$ES_HOST" --port "$ES_PORT"
  --user "$ES_USER" --password "${ES_PASSWORD:?ES_PASSWORD is required}"
  --index-name "$ES_INDEX"
  --case-type Performance768D10M
  --skip-drop-old --skip-load
  --m 32 --ef-construction 400
  --skip-search-serial --search-concurrent
  --concurrency-duration 220
)

# Top10: run two concurrency levels, c212 and c216.
vectordbbench aliyunelasticsearch "${VDB_COMMON[@]}" \
  --k 10 --num-candidates 90 \
  --use-rescore --oversample-ratio 4.6 \
  --num-concurrency 212,216 \
  --db-label cohere10m-top10-dual16-c212-c216

Test metrics

The QPS values in the results table use the higher concurrency level and are averaged over two independent 220-second VectorDBBench runs. The Serial P95 and Serial P99 values are obtained by running serial queries separately.

native_hnsw search performance

The following table lists the native_hnsw search performance results. The reproduction procedure in this whitepaper covers the Cohere10M dataset. The results for the other datasets are reference results.