Skip to content

Commit

Permalink
Add wiki_all dataset config and documentation. (#1918)
Browse files Browse the repository at this point in the history 
…et more clarify on how the dataset was generated.

Authors:
  - Corey J. Nolet (https://github.com/cjnolet)

Approvers:
  - Dante Gama Dessavre (https://github.com/dantegd)

URL: #1918
  • Loading branch information
@cjnolet
cjnolet authored Oct 25, 2023
1 parent c95d821 commit adb18b1
Show file tree
Hide file tree
Showing 5 changed files with 652 additions and 1 deletion.
7 changes: 6 additions & 1 deletion docs/source/raft_ann_benchmarks.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,8 @@ This project provides a benchmark program for various ANN search implementations
- [Running with Docker containers](#running-with-docker-containers)
- [Creating and customizing dataset configurations](#creating-and-customizing-dataset-configurations)
- [Adding a new ANN algorithm](#adding-a-new-ann-algorithm)
- [Parameter tuning guide](https://docs.rapids.ai/api/raft/nightly/ann_benchmarks_param_tuning/)
- [Wiki-all RAG/LLM Dataset](https://docs.rapids.ai/api/raft/nightly/wiki_all_dataset/)

## Installing the benchmarks

Expand Down Expand Up @@ -242,15 +244,18 @@ Configuration files already exist for the following list of the million-scale da
| `nytimes-256-angular` | 290K | 256 | 10K | Angular |
| `sift-128-euclidean` | 1M | 128 | 10K | Euclidean|
All of the datasets above contain ground test datasets with 100 neighbors. Thus `k` for these datasets must be less than or equal to 100.
All of the datasets above contain ground test datasets with 100 neighbors. Thus `k` for these datasets must be less than or equal to 100.
### End to end: large-scale benchmarks (>10M vectors)
`raft-ann-bench.get_dataset` cannot be used to download the [billion-scale datasets](ann_benchmarks_dataset.md#billion-scale)
due to their size. You should instead use our billion-scale datasets guide to download and prepare them.
All other python commands mentioned below work as intended once the
billion-scale dataset has been downloaded.
To download billion-scale datasets, visit [big-ann-benchmarks](http://big-ann-benchmarks.com/neurips21.html)
We also provide a new dataset called `wiki-all` containing 88 million 768-dimensional vectors. This dataset is meant for benchmarking a realistic retrieval-augmented generation (RAG)/LLM embedding size at scale. It also contains 1M and 10M vector subsets for smaller-scale experiments. See our [Wiki-all Dataset Guide](https://docs.rapids.ai/api/raft/nightly/wiki_all_dataset/) for more information and to download the dataset.
The steps below demonstrate how to download, install, and run benchmarks on a subset of 100M vectors from the Yandex Deep-1B dataset. Please note that datasets of this scale are recommended for GPUs with larger amounts of memory, such as the A100 or H100.
```bash
Expand Down
46 changes: 46 additions & 0 deletions docs/source/wiki_all_dataset.md
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
# Wiki-all Dataset

The `wiki-all` dataset was created to stress vector search algorithms at scale with both a large number of vectors and dimensions. The entire dataset contains 88M vectors with 768 dimensions and is meant for testing the types of vectors one would typically encounter in retrieval augmented generation (RAG) workloads. The full dataset is ~251GB in size, which is intentionally larger than the typical memory of GPUs. The massive scale is intended to promote the use of compression and efficient out-of-core methods for both indexing and search.

## Getting the dataset

The dataset is composed of all the available languages of in the [Cohere Wikipedia dataset](https://huggingface.co/datasets/Cohere/wikipedia-22-12). An [English version]( https://www.kaggle.com/datasets/jjinho/wikipedia-20230701) is also available.


The dataset is composed of English wiki texts from [Kaggle](https://www.kaggle.com/datasets/jjinho/wikipedia-20230701) and multi-lingual wiki texts from [Cohere Wikipedia](https://huggingface.co/datasets/Cohere/wikipedia-22-12).

Cohere's English Texts are older (2022) and smaller than the Kaggle English Wiki texts (2023) so the English texts have been removed from Cohere completely. The final Wiki texts include English Wiki from Kaggle and the other languages from Cohere. The English texts constitute 50% of the total text size.

To form the final dataset, the Wiki texts were chunked into 85 million 128-token pieces. For reference, Cohere chunks Wiki texts into 104-token pieces. Finally, the embeddings of each chunk were computed using the [paraphrase-multilingual-mpnet-base-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-mpnet-base-v2) embedding model. The resulting dataset is an embedding matrix of size 88 million by 768. Also included with the dataset is a query file containing 10k query vectors and a groundtruth file to evaluate nearest neighbors algorithms.

### Full dataset

A version of the dataset is made available in the binary format that can be used directly by the [raft-ann-bench](https://docs.rapids.ai/api/raft/nightly/raft_ann_benchmarks/) tool. The full 88M dataset is ~251GB and the download link below contains tarballs that have been split into multiple parts.

The following will download all 10 the parts and untar them to a `wiki_all_88M` directory:
```bash
curl -s https://data.rapids.ai/raft/datasets/wiki_all/wiki_all.tar.{00..9} | tar -xf - -C /datasets/wiki_all_88M/
```

The above has the unfortunate drawback that if the command should fail for any reason, all the parts need to be re-downloaded. The files can also be downloaded individually and then untarred to the directory. Each file is ~27GB and there are 10 of them.

```bash
curl -s https://data.rapids.ai/raft/datasets/wiki_all/wiki_all.tar.00
...
curl -s https://data.rapids.ai/raft/datasets/wiki_all/wiki_all.tar.09

cat wiki_all.tar.* | tar -xf - -C /datasets/wiki_all_88M/
```

### 1M and 10M subsets

Also available are 1M and 10M subsets of the full dataset which are 2.9GB and 29GB, respectively. These subsets also include query sets of 10k vectors and corresponding groundtruth files.

```bash
curl -s https://data.rapids.ai/raft/datasets/wiki_all_1M/wiki_all_1M.tar
curl -s https://data.rapids.ai/raft/datasets/wiki_all_10M/wiki_all_10M.tar
```

## Using the dataset

After the dataset is downloaded and extracted to the `wiki_all_88M` directory (or `wiki_all_1M`/`wiki_all_10M` depending on whether the subsets are used), the files can be used in the benchmarking tool. The dataset name is `wiki_all` (or `wiki_all_1M`/`wiki_all_10M`), and the benchmarking tool can be used by specifying the appropriate name `--dataset wiki_all_88M` in the scripts.
200 changes: 200 additions & 0 deletions python/raft-ann-bench/src/raft-ann-bench/run/conf/wiki_all_10M.json
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,200 @@
{
"dataset": {
"name": "wiki_all_10M",
"base_file": "wiki_all_10M/base.88M.fbin",
"query_file": "wiki_all_10M/queries.fbin",
"groundtruth_neighbors_file": "wiki_all_10M/groundtruth.88M.neighbors.ibin",
"distance": "euclidean"
},
"search_basic_param": {
"batch_size": 10000,
"k": 10
},
"index": [
{
"name": "hnswlib.M16.ef50",
"algo": "hnswlib",
"build_param": { "M": 16, "efConstruction": 50, "numThreads": 56 },
"file": "wiki_all_10M/hnswlib/M16.ef50",
"search_params": [
{ "ef": 10, "numThreads": 56 },
{ "ef": 20, "numThreads": 56 },
{ "ef": 40, "numThreads": 56 },
{ "ef": 60, "numThreads": 56 },
{ "ef": 80, "numThreads": 56 },
{ "ef": 120, "numThreads": 56 },
{ "ef": 200, "numThreads": 56 },
{ "ef": 400, "numThreads": 56 },
{ "ef": 600, "numThreads": 56 },
{ "ef": 800, "numThreads": 56 }
]
},
{
"name": "faiss_ivf_pq.M32-nlist16K",
"algo": "faiss_gpu_ivf_pq",
"build_param": {
"M": 32,
"nlist": 16384,
"ratio": 2
},
"file": "wiki_all_10M/faiss_ivf_pq/M32-nlist16K_ratio2",
"search_params": [
{ "nprobe": 10 },
{ "nprobe": 20 },
{ "nprobe": 30 },
{ "nprobe": 40 },
{ "nprobe": 50 },
{ "nprobe": 100 },
{ "nprobe": 200 },
{ "nprobe": 500 }
]
},
{
"name": "faiss_ivf_pq.M64-nlist16K",
"algo": "faiss_gpu_ivf_pq",
"build_param": {
"M": 64,
"nlist": 16384,
"ratio": 2
},
"file": "wiki_all_10M/faiss_ivf_pq/M64-nlist16K_ratio2",
"search_params": [
{ "nprobe": 10 },
{ "nprobe": 20 },
{ "nprobe": 30 },
{ "nprobe": 40 },
{ "nprobe": 50 },
{ "nprobe": 100 },
{ "nprobe": 200 },
{ "nprobe": 500 }
]
},
{
"name": "raft_ivf_pq.d128-nlist16K",
"algo": "raft_ivf_pq",
"build_param": {
"pq_dim": 128,
"pq_bits": 8,
"nlist": 16384,
"niter": 10,
"ratio": 10
},
"file": "wiki_all_10M/raft_ivf_pq/d128-nlist16K",
"search_params": [
{ "nprobe": 20, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 },
{ "nprobe": 30, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 },
{ "nprobe": 40, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 },
{ "nprobe": 50, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 },
{ "nprobe": 100, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 },
{ "nprobe": 200, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 },
{ "nprobe": 500, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 1 }
]
},
{
"name": "raft_ivf_pq.d64-nlist16K",
"algo": "raft_ivf_pq",
"build_param": {
"pq_dim": 64,
"pq_bits": 8,
"nlist": 16384,
"niter": 10,
"ratio": 10
},
"file": "wiki_all_10M/raft_ivf_pq/d64-nlist16K",
"search_params": [
{ "nprobe": 20, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 30, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 40, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 50, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 100, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 200, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 500, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 }
]
},
{
"name": "raft_ivf_pq.d32-nlist16K",
"algo": "raft_ivf_pq",
"build_param": {
"pq_dim": 32,
"pq_bits": 8,
"nlist": 16384,
"niter": 10,
"ratio": 10
},
"file": "wiki_all_10M/raft_ivf_pq/d32-nlist16K",
"search_params": [
{ "nprobe": 20, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 },
{ "nprobe": 30, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 },
{ "nprobe": 40, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 },
{ "nprobe": 50, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 },
{ "nprobe": 100, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 },
{ "nprobe": 200, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 },
{ "nprobe": 500, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 32 }
]
},
{
"name": "raft_ivf_pq.d32X-nlist16K",
"algo": "raft_ivf_pq",
"build_param": {
"pq_dim": 32,
"pq_bits": 8,
"nlist": 16384,
"niter": 10,
"ratio": 10
},
"file": "wiki_all_10M/raft_ivf_pq/d32-nlist16K",
"search_params": [
{ "nprobe": 20, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 30, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 40, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 50, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 100, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 200, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 500, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 16 },
{ "nprobe": 30, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 8 },
{ "nprobe": 40, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 8 },
{ "nprobe": 50, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 8 },
{ "nprobe": 100, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 8 },
{ "nprobe": 200, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 8 },
{ "nprobe": 500, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 8 },
{ "nprobe": 30, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 40, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 50, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 100, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 200, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 },
{ "nprobe": 500, "internalDistanceDtype": "half", "smemLutDtype": "half", "refine_ratio": 4 }

]
},
{
"name": "raft_cagra.dim32.multi_cta",
"algo": "raft_cagra",
"build_param": { "graph_degree": 32, "intermediate_graph_degree": 48 },
"file": "wiki_all_10M/raft_cagra/dim32.ibin",
"search_params": [
{ "itopk": 32, "search_width": 1, "max_iterations": 0, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 1, "max_iterations": 32, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 1, "max_iterations": 36, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 1, "max_iterations": 40, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 1, "max_iterations": 44, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 1, "max_iterations": 48, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 2, "max_iterations": 16, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 2, "max_iterations": 24, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 2, "max_iterations": 26, "algo": "multi_cta" },
{ "itopk": 32, "search_width": 2, "max_iterations": 32, "algo": "multi_cta" },
{ "itopk": 64, "search_width": 4, "max_iterations": 16, "algo": "multi_cta" },
{ "itopk": 64, "search_width": 1, "max_iterations": 64, "algo": "multi_cta" },
{ "itopk": 96, "search_width": 2, "max_iterations": 48, "algo": "multi_cta" },
{ "itopk": 128, "search_width": 8, "max_iterations": 16, "algo": "multi_cta" },
{ "itopk": 128, "search_width": 2, "max_iterations": 64, "algo": "multi_cta" },
{ "itopk": 192, "search_width": 8, "max_iterations": 24, "algo": "multi_cta" },
{ "itopk": 192, "search_width": 2, "max_iterations": 96, "algo": "multi_cta" },
{ "itopk": 256, "search_width": 8, "max_iterations": 32, "algo": "multi_cta" },
{ "itopk": 384, "search_width": 8, "max_iterations": 48, "algo": "multi_cta" },
{ "itopk": 512, "search_width": 8, "max_iterations": 64, "algo": "multi_cta" }
]
}

]
}

Loading

0 comments on commit adb18b1

Please sign in to comment.