Skip to content

Commit

Permalink
Add vectorsearch training workload (#333) (#345)
Browse files Browse the repository at this point in the history 
* Add vectorsearch training workload



* Addressed Vijay feedback and ignores error if model DNE



* Added documentation to VS readme



---------


(cherry picked from commit 29d9715)

Signed-off-by: Finn Roblin <[email protected]>
Signed-off-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
  • Loading branch information
@opensearch-trigger-bot @github-actions
opensearch-trigger-bot[bot] and github-actions[bot] authored Jul 18, 2024
1 parent c893ef7 commit 7dc61ad
Show file tree
Hide file tree
Showing 10 changed files with 444 additions and 2 deletions.
142 changes: 142 additions & 0 deletions vectorsearch/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -189,6 +189,148 @@ either in-memory, or an external OpenSearch cluster.
| error rate | prod-queries | 0 | % |
---------------------------------
[INFO] SUCCESS (took 119 seconds)
---------------------------------
```

### Train Test

This procedure benchmarks approximate k-NN search algorithms that require a training step. For example, the FAISS IVF requires a training step to cluster vectors. Then search can be performed against a smaller number of cluster centroids instead of the entire dataset.

#### Parameters

This workload allows the following parameters to be specified using `--workload-params`:

| Name | Description |
|-----------------------------------------|----------------------------------------------------------------------------------------------|
| target_index_name | Name of index to add vectors to |
| target_field_name | Name of field to add vectors to |
| target_index_body | Path to target index definition |
| target_index_primary_shards | Target index primary shards |
| target_index_replica_shards | Target index replica shards |
| target_index_dimension | Dimension of target index |
| target_index_space_type | Target index space type |
| target_index_bulk_size | Target index bulk size |
| target_index_bulk_index_data_set_format | Format of vector data set |
| target_index_bulk_index_data_set_path | Path to vector data set |
| target_index_bulk_index_data_set_corpus | Corpus name to vector data set |
| target_index_bulk_index_clients | Clients to be used for bulk ingestion (must be divisor of data set size) |
| target_index_max_num_segments | Number of segments to merge target index down to before beginning search |
| target_index_force_merge_timeout | Timeout for of force merge requests in seconds |
| train_index_name | Name of index for training |
| train_field_name | Name of field for training |
| train_method_engine | Engine for training (e.g "faiss") |
| train_index_body | Path to train index definition |
| train_index_primary_shards | Train index primary shards |
| train_index_replica_shards | Train index replica shards |
| train_index_bulk_size | Bulk size for train index |
| train_index_bulk_index_data_set_format | Format of training data set |
| train_index_bulk_index_data_set_path | Path to training data set |
| train_index_bulk_indexing_clients | Clients to be used for bulk indexing |
| train_index_num_vectors | Number of vectors in the training index |
| train_model_id | ID of the training model |
| train_operation_retries | Number of retries for querying training operation to see if complete |
| train_operation_poll_period | Poll period for querying training operation in seconds |
| train_search_size | Number of results per [scroll query](http://opensearch.org/docs/latest/api-reference/scroll/)|
| hnsw_ef_search | HNSW ef search parameter |
| hnsw_ef_construction | HNSW ef construction parameter |
| id_field_name | Name of field that will be used to identify documents in an index |
| hnsw_m | HNSW m parameter |
| query_k | The number of neighbors to return for the search |
| query_data_set_format | Format of vector data set for queries |
| query_data_set_path | Path to vector data set for queries |
| query_count | Number of queries for search operation |
| query_body | Json properties that will be merged with search body |
| search_clients | Number of clients to use for running queries |

#### Sample Output

The output of a sample test run is provided below. Metrics are captured in the result's data store as usual, and this can be configured to be
either in-memory, or an external OpenSearch cluster.

```
------------------------------------------------------
_______ __ _____
/ ____(_)___ ____ _/ / / ___/_________ ________
/ /_ / / __ \/ __ `/ / \__ \/ ___/ __ \/ ___/ _ \
/ __/ / / / / / /_/ / / ___/ / /__/ /_/ / / / __/
/_/ /_/_/ /_/\__,_/_/ /____/\___/\____/_/ \___/
------------------------------------------------------
| Metric | Task | Value | Unit |
|---------------------------------------------------------------:|-------------------:|------------:|-------:|
| Cumulative indexing time of primary shards | | 0.00946667 | min |
| Min cumulative indexing time across primary shards | | 0 | min |
| Median cumulative indexing time across primary shards | | 0.00298333 | min |
| Max cumulative indexing time across primary shards | | 0.00336667 | min |
| Cumulative indexing throttle time of primary shards | | 0 | min |
| Min cumulative indexing throttle time across primary shards | | 0 | min |
| Median cumulative indexing throttle time across primary shards | | 0 | min |
| Max cumulative indexing throttle time across primary shards | | 0 | min |
| Cumulative merge time of primary shards | | 0 | min |
| Cumulative merge count of primary shards | | 0 | |
| Min cumulative merge time across primary shards | | 0 | min |
| Median cumulative merge time across primary shards | | 0 | min |
| Max cumulative merge time across primary shards | | 0 | min |
| Cumulative merge throttle time of primary shards | | 0 | min |
| Min cumulative merge throttle time across primary shards | | 0 | min |
| Median cumulative merge throttle time across primary shards | | 0 | min |
| Max cumulative merge throttle time across primary shards | | 0 | min |
| Cumulative refresh time of primary shards | | 0.00861667 | min |
| Cumulative refresh count of primary shards | | 33 | |
| Min cumulative refresh time across primary shards | | 0 | min |
| Median cumulative refresh time across primary shards | | 0.00268333 | min |
| Max cumulative refresh time across primary shards | | 0.00291667 | min |
| Cumulative flush time of primary shards | | 0.000183333 | min |
| Cumulative flush count of primary shards | | 2 | |
| Min cumulative flush time across primary shards | | 0 | min |
| Median cumulative flush time across primary shards | | 0 | min |
| Max cumulative flush time across primary shards | | 0.000183333 | min |
| Total Young Gen GC time | | 0.075 | s |
| Total Young Gen GC count | | 17 | |
| Total Old Gen GC time | | 0 | s |
| Total Old Gen GC count | | 0 | |
| Store size | | 0.00869293 | GB |
| Translog size | | 2.56114e-07 | GB |
| Heap used for segments | | 0 | MB |
| Heap used for doc values | | 0 | MB |
| Heap used for terms | | 0 | MB |
| Heap used for norms | | 0 | MB |
| Heap used for points | | 0 | MB |
| Heap used for stored fields | | 0 | MB |
| Segment count | | 9 | |
| Min Throughput | custom-vector-bulk | 25527 | docs/s |
| Mean Throughput | custom-vector-bulk | 25527 | docs/s |
| Median Throughput | custom-vector-bulk | 25527 | docs/s |
| Max Throughput | custom-vector-bulk | 25527 | docs/s |
| 50th percentile latency | custom-vector-bulk | 36.3095 | ms |
| 90th percentile latency | custom-vector-bulk | 52.2662 | ms |
| 100th percentile latency | custom-vector-bulk | 68.6513 | ms |
| 50th percentile service time | custom-vector-bulk | 36.3095 | ms |
| 90th percentile service time | custom-vector-bulk | 52.2662 | ms |
| 100th percentile service time | custom-vector-bulk | 68.6513 | ms |
| error rate | custom-vector-bulk | 0 | % |
| Min Throughput | prod-queries | 211.26 | ops/s |
| Mean Throughput | prod-queries | 213.85 | ops/s |
| Median Throughput | prod-queries | 213.48 | ops/s |
| Max Throughput | prod-queries | 216.49 | ops/s |
| 50th percentile latency | prod-queries | 3.43393 | ms |
| 90th percentile latency | prod-queries | 4.01881 | ms |
| 99th percentile latency | prod-queries | 5.56238 | ms |
| 99.9th percentile latency | prod-queries | 9.95666 | ms |
| 99.99th percentile latency | prod-queries | 39.7922 | ms |
| 100th percentile latency | prod-queries | 62.415 | ms |
| 50th percentile service time | prod-queries | 3.43405 | ms |
| 90th percentile service time | prod-queries | 4.0191 | ms |
| 99th percentile service time | prod-queries | 5.56316 | ms |
| 99.9th percentile service time | prod-queries | 9.95666 | ms |
| 99.99th percentile service time | prod-queries | 39.7922 | ms |
| 100th percentile service time | prod-queries | 62.415 | ms |
| error rate | prod-queries | 0 | % |
---------------------------------
[INFO] SUCCESS (took 119 seconds)
---------------------------------
Expand Down
21 changes: 21 additions & 0 deletions vectorsearch/indices/train-index.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,21 @@
{
"settings": {
"index": {
{%- if train_index_primary_shards is defined and train_index_primary_shards %}
"number_of_shards": {{ train_index_primary_shards }}
{%- endif %}
{%- if train_index_replica_shards is defined %}
,"number_of_replicas": {{ train_index_replica_shards }}
{%- endif %}
}
},
"mappings": {
"properties": {
"{{ train_field_name }}": {
"type": "knn_vector",
"dimension": {{ target_index_dimension }}
}
}
}
}

12 changes: 11 additions & 1 deletion vectorsearch/operations/default.json
View file
Original file line number Diff line number Diff line change
Expand Up @@ -18,4 +18,14 @@
"operation-type": "refresh",
"retries": 100,
"index": "{{ target_index_name | default('target_index') }}"
}
},
{
"name": "refresh-train-index",
"operation-type": "refresh",
"retries": 100,
"index": "{{ train_index_name | default('train_index') }}"
},
{
"name": "delete-model",
"operation-type": "delete-knn-model"
}
48 changes: 48 additions & 0 deletions vectorsearch/params/train/train-faiss-sift-128-l2-pq.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,48 @@
{
"target_index_name": "target_index",
"target_field_name": "target_field",
"target_index_body": "indices/faiss-index.json",
"target_index_primary_shards": 1,
"target_index_dimension": 128,
"target_index_space_type": "l2",
"target_index_bulk_size": 100,
"target_index_bulk_index_data_set_format": "hdf5",
"target_index_bulk_index_data_set_path": "/tmp/sift-128-euclidean.hdf5",
"target_index_bulk_indexing_clients": 10,

"train_index_name": "train_index",
"train_field_name": "train_field",
"train_method_engine": "faiss",
"train_index_body": "indices/train-index.json",
"train_index_primary_shards": 1,
"train_index_replica_shards": 1,

"train_index_bulk_size": 100,
"train_index_bulk_index_data_set_format": "hdf5",
"train_index_bulk_index_data_set_path": "/tmp/sift-128-euclidean.hdf5",
"train_index_bulk_indexing_clients": 10,
"train_index_num_vectors": 1000,

"train_model_id": "test-model",
"train_operation_retries": 100,
"train_operation_poll_period": 0.5,
"train_search_size": 10000,

"encoder": "pq",
"faiss_encoder_code_size": 2,
"faiss_encoder_m": 4,

"target_index_max_num_segments": 1,
"target_index_force_merge_timeout": 300,
"hnsw_ef_search": 100,
"hnsw_ef_construction": 100,
"query_k": 100,
"query_body": {
"docvalue_fields" : ["_id"],
"stored_fields" : "_none_"
},

"query_data_set_format": "hdf5",
"query_data_set_path":"/tmp/sift-128-euclidean.hdf5",
"query_count": 100
}
48 changes: 48 additions & 0 deletions vectorsearch/params/train/train-faiss-sift-128-l2-sq.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,48 @@
{
"target_index_name": "target_index",
"target_field_name": "target_field",
"target_index_body": "indices/faiss-index.json",
"target_index_primary_shards": 1,
"target_index_dimension": 128,
"target_index_space_type": "l2",
"target_index_bulk_size": 100,
"target_index_bulk_index_data_set_format": "hdf5",
"target_index_bulk_index_data_set_path": "/tmp/sift-128-euclidean.hdf5",
"target_index_bulk_indexing_clients": 10,

"train_index_name": "train_index",
"train_field_name": "train_field",
"train_method_engine": "faiss",
"train_index_body": "indices/train-index.json",
"train_index_primary_shards": 1,
"train_index_replica_shards": 1,

"train_index_bulk_size": 100,
"train_index_bulk_index_data_set_format": "hdf5",
"train_index_bulk_index_data_set_path": "/tmp/sift-128-euclidean.hdf5",
"train_index_bulk_indexing_clients": 10,
"train_index_num_vectors": 1000,

"train_model_id": "test-model",
"train_operation_retries": 100,
"train_operation_poll_period": 0.5,
"train_search_size": 10000,

"encoder": "sq",
"faiss_encoder_type": "fp16",
"faiss_encoder_clip": false,

"target_index_max_num_segments": 1,
"target_index_force_merge_timeout": 300,
"hnsw_ef_search": 100,
"hnsw_ef_construction": 100,
"query_k": 100,
"query_body": {
"docvalue_fields" : ["_id"],
"stored_fields" : "_none_"
},

"query_data_set_format": "hdf5",
"query_data_set_path":"/tmp/sift-128-euclidean.hdf5",
"query_count": 100
}
Loading

0 comments on commit 7dc61ad

Please sign in to comment.