Building vector indexes remotely using GPUs

Introduced 3.0

OpenSearch supports building vector indexes using a GPU-accelerated remote index build service. Using GPUs dramatically reduces index build times and decreases costs. For benchmarking results, see this blog post.

Supported configurations

The remote index build service supports Faiss indexes that use the hnsw method. For these indexes, the service supports the following vector types:

• The default 32-bit floating-point (FP32) vectors
• 16-bit floating-point (FP16), byte, and binary vectors, at all compression levels (2x, 8x, 16x, and 32x)
• half_float vectors, at the 1x and 16x compression levels

Vectors quantized using the bf16 encoder type are not supported, and indexes using that encoder are always built locally.

Prerequisites

Before configuring the remote index build settings, ensure you fulfill the following prerequisites. For more information about updating dynamic settings, see Dynamic settings.

Step 1: Enable the remote index build service

OpenSearch builds an index remotely only when both the cluster-level knn.remote_index_build.enabled setting and the index-level index.knn.remote_index_build.enabled setting are true. The cluster-level setting is false by default, so enable it for the cluster:

PUT /_cluster/settings
{
  "persistent": {
    "knn.remote_index_build.enabled": true
  }
}

The index-level setting is true by default, so set it to false only to exclude an individual index from remote index building. For descriptions of both settings, see Remote index build settings.

Step 2: Create and register the remote vector repository

The remote vector repository acts as an intermediate object store between the OpenSearch cluster and the remote build service. The cluster uploads vectors and document IDs to the repository. The remote build service retrieves the data, builds the index externally, and uploads the completed result back to the repository.

To create and register the repository, follow the steps in Register repository. Then set the knn.remote_index_build.repository dynamic setting to the name of the registered repository.

The remote build service only supports Amazon Simple Storage Service (Amazon S3) repositories.

Step 3: Set up a remote vector index builder

Configure the remote endpoint in the k-NN settings by setting knn.remote_index_build.service.endpoint to a running remote vector index builder instance. For instructions on setting up the remote service, see the user guide.

Configuring remote index build settings

The remote index build service supports several additional, optional settings. For information about configuring any remaining remote index build settings, see Remote index build settings.

Using the remote index build service

Once the remote index build service is configured, any segment flush and merge operations that meet the following requirements will transparently use the GPU build path:

• The index is using one of the supported configurations.
• The segment size is greater than index.knn.remote_index_build.size.min and less than knn.remote_index_build.size.max.

You can monitor remote index build tasks by calling the k-NN Stats API and reviewing the remote index build statistics.