qdrant/lib/segment/benches/sparse_index_build.rs

#[cfg(not(target_os = "windows"))]
mod prof;

use std::borrow::Cow;
use std::sync::Arc;
use std::sync::atomic::AtomicBool;

use atomic_refcell::AtomicRefCell;
use common::counter::hardware_counter::HardwareCounterCell;
use common::types::PointOffsetType;
use criterion::{Criterion, criterion_group, criterion_main};
use half::f16;
use rand::SeedableRng;
use rand::rngs::StdRng;
use segment::common::rocksdb_wrapper::{DB_VECTOR_CF, open_db};
use segment::fixtures::payload_context_fixture::create_id_tracker_fixture;
use segment::index::VectorIndex;
use segment::index::sparse_index::sparse_index_config::{SparseIndexConfig, SparseIndexType};
use segment::index::sparse_index::sparse_vector_index::{
    SparseVectorIndex, SparseVectorIndexOpenArgs,
};
use segment::index::struct_payload_index::StructPayloadIndex;
use segment::payload_storage::in_memory_payload_storage::InMemoryPayloadStorage;
use segment::types::VectorStorageDatatype;
use segment::vector_storage::VectorStorage;
use segment::vector_storage::sparse::simple_sparse_vector_storage::open_simple_sparse_vector_storage;
use sparse::common::sparse_vector_fixture::random_sparse_vector;
use sparse::index::inverted_index::InvertedIndex;
use sparse::index::inverted_index::inverted_index_compressed_mmap::InvertedIndexCompressedMmap;
use sparse::index::inverted_index::inverted_index_mmap::InvertedIndexMmap;
use sparse::index::inverted_index::inverted_index_ram::InvertedIndexRam;
use tempfile::Builder;

const NUM_VECTORS: usize = 10_000;
const MAX_SPARSE_DIM: usize = 1_000;

fn sparse_vector_index_build_benchmark(c: &mut Criterion) {
    let mut group = c.benchmark_group("sparse-vector-build-group");

    let stopped = AtomicBool::new(false);
    let mut rnd = StdRng::seed_from_u64(42);

    let payload_dir = Builder::new().prefix("payload_dir").tempdir().unwrap();
    let storage_dir = Builder::new().prefix("storage_dir").tempdir().unwrap();
    let index_dir = Builder::new().prefix("index_dir").tempdir().unwrap();

    // setup
    let id_tracker = Arc::new(AtomicRefCell::new(create_id_tracker_fixture(NUM_VECTORS)));
    let payload_storage = InMemoryPayloadStorage::default();
    let wrapped_payload_storage = Arc::new(AtomicRefCell::new(payload_storage.into()));
    let payload_index = StructPayloadIndex::open(
        wrapped_payload_storage,
        id_tracker.clone(),
        std::collections::HashMap::new(),
        payload_dir.path(),
        true,
        true,
    )
    .unwrap();
    let wrapped_payload_index = Arc::new(AtomicRefCell::new(payload_index));

    let db = open_db(storage_dir.path(), &[DB_VECTOR_CF]).unwrap();
    let mut vector_storage = open_simple_sparse_vector_storage(db, DB_VECTOR_CF, &stopped).unwrap();

    let hw_counter = HardwareCounterCell::new();

    // add points to storage only once
    for idx in 0..NUM_VECTORS {
        let vec = &random_sparse_vector(&mut rnd, MAX_SPARSE_DIM);
        vector_storage
            .insert_vector(idx as PointOffsetType, vec.into(), &hw_counter)
            .unwrap();
    }

    // save index config to disk
    let index_config = SparseIndexConfig::new(
        Some(10_000),
        SparseIndexType::MutableRam,
        Some(VectorStorageDatatype::Float32),
    );

    let vector_storage = Arc::new(AtomicRefCell::new(vector_storage));

    // intent: measure in-memory build time from storage
    group.bench_function("build-ram-index", |b| {
        b.iter(|| {
            let sparse_vector_index: SparseVectorIndex<InvertedIndexRam> =
                SparseVectorIndex::open(SparseVectorIndexOpenArgs {
                    config: index_config,
                    id_tracker: id_tracker.clone(),
                    vector_storage: vector_storage.clone(),
                    payload_index: wrapped_payload_index.clone(),
                    path: index_dir.path(),
                    stopped: &stopped,
                    tick_progress: || (),
                })
                .unwrap();
            assert_eq!(sparse_vector_index.indexed_vector_count(), NUM_VECTORS);
        })
    });

    // build once to reuse in mmap conversion benchmark
    let sparse_vector_index: SparseVectorIndex<InvertedIndexRam> =
        SparseVectorIndex::open(SparseVectorIndexOpenArgs {
            config: index_config,
            id_tracker,
            vector_storage,
            payload_index: wrapped_payload_index,
            path: index_dir.path(),
            stopped: &stopped,
            tick_progress: || (),
        })
        .unwrap();

    // intent: measure mmap conversion time
    group.bench_function("convert-mmap-index", |b| {
        b.iter(|| {
            let mmap_index_dir = Builder::new().prefix("mmap_index_dir").tempdir().unwrap();
            let mmap_inverted_index = InvertedIndexMmap::from_ram_index(
                Cow::Borrowed(sparse_vector_index.inverted_index()),
                &mmap_index_dir,
            )
            .unwrap();
            assert_eq!(mmap_inverted_index.vector_count(), NUM_VECTORS);
        })
    });

    group.bench_function("convert-mmap-index-f32", |b| {
        b.iter(|| {
            let mmap_index_dir = Builder::new().prefix("mmap_index_dir").tempdir().unwrap();
            let mmap_inverted_index = InvertedIndexCompressedMmap::<f32>::from_ram_index(
                Cow::Borrowed(sparse_vector_index.inverted_index()),
                &mmap_index_dir,
            )
            .unwrap();
            assert_eq!(mmap_inverted_index.vector_count(), NUM_VECTORS);
        })
    });

    group.bench_function("convert-mmap-index-f16", |b| {
        b.iter(|| {
            let mmap_index_dir = Builder::new().prefix("mmap_index_dir").tempdir().unwrap();
            let mmap_inverted_index = InvertedIndexCompressedMmap::<f16>::from_ram_index(
                Cow::Borrowed(sparse_vector_index.inverted_index()),
                &mmap_index_dir,
            )
            .unwrap();
            assert_eq!(mmap_inverted_index.vector_count(), NUM_VECTORS);
        })
    });

    group.finish();
}

#[cfg(not(target_os = "windows"))]
criterion_group! {
    name = benches;
    config = Criterion::default().with_profiler(prof::FlamegraphProfiler::new(100));
    targets = sparse_vector_index_build_benchmark
}

#[cfg(target_os = "windows")]
criterion_group! {
    name = benches;
    config = Criterion::default();
    targets = sparse_vector_index_build_benchmark,
}

criterion_main!(benches);