Skip to content

Optimize transfer/compute overlap in out-of-core KMeans #2538

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
viclafargue wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Optimize transfer/compute overlap in out-of-core KMeans #2538

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
95e3729
Optimize out-of-core KMeans overlap
viclafargue Sep 1, 2026
9c16817
Refine out-of-core KMeans batch pipelining
viclafargue Sep 1, 2026
696e22d
Merge branch 'main' into ooc-kmeans-overlap
viclafargue Sep 1, 2026
4b8a5c5
multi-GPU transfer/compute overlap
viclafargue Sep 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
212 changes: 147 additions & 65 deletions cpp/src/cluster/detail/kmeans.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@
#pragma once

#include "../../core/nvtx.hpp"
#include "../../neighbors/detail/ann_utils.cuh"
#include "kmeans_batch_loader.cuh"
#include "kmeans_common.cuh"

#include <cuvs/cluster/kmeans.hpp>
Expand All @@ -24,6 +24,8 @@
#include <raft/core/pinned_mdarray.hpp>
#include <raft/core/pinned_mdspan.hpp>
#include <raft/core/resource/cuda_stream.hpp>
#include <raft/core/resource/cuda_stream_pool.hpp>
#include <raft/core/resource/device_memory_resource.hpp>
#include <raft/core/resource/thrust_policy.hpp>
#include <raft/core/resources.hpp>
#include <raft/linalg/map.cuh>
Expand All @@ -39,6 +41,7 @@
#include <raft/util/cuda_utils.cuh>
#include <raft/util/cudart_utils.hpp>

#include <rmm/cuda_stream.hpp>
#include <rmm/device_scalar.hpp>
#include <rmm/device_uvector.hpp>

Expand Down Expand Up @@ -686,25 +689,56 @@ void kmeans_fit(

auto minClusterAndDistance = raft::make_device_vector<raft::KeyValuePair<IndexT, DataT>, IndexT>(
handle, device_buffer_samples);
auto L2NormBatch = raft::make_device_vector<DataT, IndexT>(handle, device_buffer_samples);
const IndexT l2_norm_size = data_on_device ? n_samples : device_buffer_samples;
auto L2NormBatch = raft::make_device_vector<DataT, IndexT>(handle, l2_norm_size);
auto batch_weights_buf = raft::make_device_vector<DataT, IndexT>(handle, device_buffer_samples);
rmm::device_uvector<DataT> L2NormBuf_OR_DistBuf(0, stream);

auto centroid_sums = raft::make_device_matrix<DataT, IndexT>(handle, n_clusters, n_features);
auto weight_per_cluster = raft::make_device_vector<DataT, IndexT>(handle, n_clusters);
auto clustering_cost = raft::make_device_scalar<DataT>(handle, DataT{0});
auto batch_inertia = raft::make_device_scalar<DataT>(handle, DataT{0});
auto batch_cost = raft::make_device_scalar<DataT>(handle, DataT{0});
auto h_inertia = raft::make_pinned_scalar<DataT>(handle, DataT{0});

rmm::device_uvector<char> batch_workspace(device_buffer_samples, stream);

auto data_batches = cuvs::spatial::knn::detail::utils::make_batch_load_iterator<DataT>(
handle, X.data_handle(), n_samples, n_features, device_buffer_samples, stream);
auto batch_mr = data_on_device ? raft::resource::get_workspace_resource_ref(handle)
: raft::resource::get_large_workspace_resource_ref(handle);
auto batch_copy_stream = raft::resource::get_cuda_stream(handle);
std::optional<rmm::cuda_stream> owned_batch_copy_stream;
if constexpr (!data_on_device) {
// Host KMeans must pipeline even when a minimal caller (including the Python binding) did not
// install a RAFT stream pool. Keep the stream owner before the iterators so their destructors
// drain all event dependencies before the stream itself is destroyed.
if (handle.has_resource_factory(raft::resource::resource_type::CUDA_STREAM_POOL) &&
raft::resource::get_stream_pool_size(handle) >= 1) {
batch_copy_stream = raft::resource::get_stream_from_stream_pool(handle);
} else {
owned_batch_copy_stream.emplace(rmm::cuda_stream::flags::non_blocking);

@tarang-jain tarang-jain Sep 3, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Lets not create a prefetch stream on our own. It is solely controlled by the raft resource. The resource should give full control of concurrent streams to the user.

batch_copy_stream = owned_batch_copy_stream->view();
}
}

kmeans_batch_loader<DataT, IndexT, data_on_device> data_batches(handle,
X.data_handle(),
n_samples,
n_features,
device_buffer_samples,
batch_copy_stream,
batch_mr);
// Host-path weight batches: only materialized when weights are provided and
// the data resides on host
std::optional<cuvs::spatial::knn::detail::utils::batch_load_iterator_dyn<DataT>> weight_batches;
std::optional<kmeans_batch_loader<DataT, IndexT, false>> weight_batches;
if constexpr (!data_on_device) {
if (weight_ptr != nullptr) {
weight_batches = cuvs::spatial::knn::detail::utils::make_batch_load_iterator<DataT>(
handle, weight_ptr, n_samples, IndexT{1}, device_buffer_samples, stream);
weight_batches.emplace(handle,
weight_ptr,
n_samples,
IndexT{1},
device_buffer_samples,
batch_copy_stream,
batch_mr);
} else {
raft::matrix::fill(handle, batch_weights_buf.view(), DataT{1});
}
Expand Down Expand Up @@ -758,6 +792,18 @@ void kmeans_fit(
}
};

auto prefetch_batch = [&](std::size_t batch_pos) {
(void)data_batches.prefetch(batch_pos);
if (weight_batches.has_value()) { (void)weight_batches->prefetch(batch_pos); }
};

bool input_pipeline_started = false;
auto start_input_pipeline = [&] {
if (input_pipeline_started) { return; }
if (data_batches.num_batches() > 0) { prefetch_batch(0); }
input_pipeline_started = true;
};

RAFT_LOG_DEBUG(
"KMeans.fit: n_samples=%zu, n_features=%zu, n_clusters=%d, device_buffer_samples=%zu",
static_cast<size_t>(n_samples),
Expand All @@ -767,10 +813,6 @@ void kmeans_fit(

bool need_compute_norms = metric == cuvs::distance::DistanceType::L2Expanded ||
metric == cuvs::distance::DistanceType::L2SqrtExpanded;
auto h_norm_cache = raft::make_pinned_vector<DataT, IndexT>(
handle, (need_compute_norms && !data_on_device) ? n_samples : 0);
bool norms_cached = false;

auto compute_batch_norms = [&](const DataT* batch_ptr, IndexT batch_size) {
auto batch_view =
raft::make_device_matrix_view<const DataT, IndexT>(batch_ptr, batch_size, n_features);
Expand Down Expand Up @@ -830,53 +872,44 @@ void kmeans_fit(
raft::matrix::fill(handle, weight_per_cluster.view(), DataT{0});
raft::matrix::fill(handle, clustering_cost.view(), DataT{0});

// Complete iteration setup before starting the cold pipeline, so no potentially blocking
// CUDA setup remains between the first transfer and its first consumer.
start_input_pipeline();

auto centroids_const = raft::make_device_matrix_view<const DataT, IndexT>(
cur_centroids_ptr, n_clusters, n_features);
auto new_centroids_view =
raft::make_device_matrix_view<DataT, IndexT>(new_centroids_ptr, n_clusters, n_features);

data_batches.reset();
using wt_iter_t = cuvs::spatial::knn::detail::utils::batch_load_iterator_dyn<DataT>;
std::optional<wt_iter_t> wt_it;
if (weight_batches.has_value()) {
weight_batches->reset();
wt_it = weight_batches->begin();
}
for (const auto& data_batch : data_batches) {
IndexT cur_batch_size = static_cast<IndexT>(data_batch.size());
const DataT* wt_data = nullptr;
if (wt_it.has_value()) {
wt_data = (**wt_it).data();
++(*wt_it);
for (std::size_t batch_pos = 0; batch_pos < data_batches.num_batches(); ++batch_pos) {
const auto data_batch = data_batches.acquire(batch_pos);
std::optional<kmeans_batch<DataT>> weight_batch;
if (weight_batches.has_value()) {
weight_batch.emplace(weight_batches->acquire(batch_pos));
}

IndexT cur_batch_size = static_cast<IndexT>(data_batch.size());
const DataT* wt_data = weight_batch.has_value() ? weight_batch->data() : nullptr;

auto batch_data_view = raft::make_device_matrix_view<const DataT, IndexT>(
data_batch.data(), cur_batch_size, n_features);
auto batch_weights_view =
cur_batch_weights(static_cast<IndexT>(data_batch.offset()), wt_data, cur_batch_size);

auto minCAD_view = raft::make_device_vector_view<raft::KeyValuePair<IndexT, DataT>, IndexT>(
minClusterAndDistance.data_handle(), cur_batch_size);

if constexpr (!data_on_device) {
if (need_compute_norms) {
if (!norms_cached) {
compute_batch_norms(data_batch.data(), cur_batch_size);
raft::copy(h_norm_cache.data_handle() + data_batch.offset(),
L2NormBatch.data_handle(),
cur_batch_size,
stream);
} else {
raft::copy(L2NormBatch.data_handle(),
h_norm_cache.data_handle() + data_batch.offset(),
cur_batch_size,
stream);
}
}
if (need_compute_norms) { compute_batch_norms(data_batch.data(), cur_batch_size); }
}

// An already-full pipeline makes this a no-op. During cold fill, submit the first real
// consumer before making the second H2D eligible, so CUDA can dispatch both at batch-ready.
prefetch_batch((batch_pos + 1) % data_batches.num_batches());

const auto l2_norm_offset =
data_on_device ? static_cast<IndexT>(data_batch.offset()) : IndexT{0};
auto l2_const_view = raft::make_device_vector_view<const DataT, IndexT>(
L2NormBatch.data_handle(), cur_batch_size);
L2NormBatch.data_handle() + l2_norm_offset, cur_batch_size);

process_batch<DataT, IndexT>(handle,
batch_data_view,
Expand All @@ -892,9 +925,15 @@ void kmeans_fit(
centroid_sums.view(),
weight_per_cluster.view(),
clustering_cost.view(),
batch_workspace);
batch_workspace,
batch_cost.view());

// The slot is reusable only after every batch consumer above has been submitted. Refill it
// with the batch two positions ahead; modulo arithmetic naturally crosses pass boundaries.
const auto next_batch_pos = (batch_pos + 2) % data_batches.num_batches();
data_batches.recycle(data_batch, next_batch_pos);
if (weight_batch.has_value()) { weight_batches->recycle(*weight_batch, next_batch_pos); }
}
if (need_compute_norms) { norms_cached = true; }

finalize_centroids<DataT, IndexT>(handle,
raft::make_const_mdspan(centroid_sums.view()),
Expand Down Expand Up @@ -927,46 +966,89 @@ void kmeans_fit(
raft::copy(handle,
raft::make_pinned_scalar_view(h_done_flag.data_handle()),
raft::make_device_scalar_view<const int>(d_done_flag.data_handle()));
// The next pass's first two input batches are already in flight. The compute stream still
// serializes centroid finalization and convergence before it can consume them.
}

{
auto centroids_const = raft::make_device_matrix_view<const DataT, IndexT>(
cur_centroids_ptr, n_clusters, n_features);

iter_inertia = DataT{0};
data_batches.reset();
using wt_iter_t = cuvs::spatial::knn::detail::utils::batch_load_iterator_dyn<DataT>;
std::optional<wt_iter_t> wt_it;
if (weight_batches.has_value()) {
weight_batches->reset();
wt_it = weight_batches->begin();
}
for (const auto& data_batch : data_batches) {
IndexT cur_batch_size = static_cast<IndexT>(data_batch.size());
const DataT* wt_data = nullptr;
if (wt_it.has_value()) {
wt_data = (**wt_it).data();
++(*wt_it);
raft::matrix::fill(handle, clustering_cost.view(), DataT{0});
start_input_pipeline();
for (std::size_t batch_pos = 0; batch_pos < data_batches.num_batches(); ++batch_pos) {
const auto data_batch = data_batches.acquire(batch_pos);
std::optional<kmeans_batch<DataT>> weight_batch;
if (weight_batches.has_value()) {
weight_batch.emplace(weight_batches->acquire(batch_pos));
}

IndexT cur_batch_size = static_cast<IndexT>(data_batch.size());
const DataT* wt_data = weight_batch.has_value() ? weight_batch->data() : nullptr;

auto batch_data_view = raft::make_device_matrix_view<const DataT, IndexT>(
data_batch.data(), cur_batch_size, n_features);

std::optional<raft::device_vector_view<const DataT, IndexT>> batch_sw = std::nullopt;
if (weight_ptr != nullptr) {
batch_sw =
cur_batch_weights(static_cast<IndexT>(data_batch.offset()), wt_data, cur_batch_size);
}
compute_batch_norms(data_batch.data(), cur_batch_size);
if (batch_pos + 1 < data_batches.num_batches() || seed_iter + 1 < n_init) {
prefetch_batch((batch_pos + 1) % data_batches.num_batches());
}

DataT batch_cost = DataT{0};
cuvs::cluster::kmeans::cluster_cost(handle,
batch_data_view,
centroids_const,
raft::make_host_scalar_view(&batch_cost),
batch_sw);

iter_inertia += batch_cost;
auto l2_norm_view = raft::make_device_vector_view<const DataT, IndexT>(
L2NormBatch.data_handle(), cur_batch_size);
auto min_cad_view =
raft::make_device_vector_view<raft::KeyValuePair<IndexT, DataT>, IndexT>(
minClusterAndDistance.data_handle(), cur_batch_size);

minClusterAndDistanceCompute<DataT, IndexT>(handle,
batch_data_view,
centroids_const,
min_cad_view,
l2_norm_view,
L2NormBuf_OR_DistBuf,
cuvs::distance::DistanceType::L2Expanded,
iter_params.batch_samples,
iter_params.batch_centroids,
ws);
if (batch_sw.has_value()) {
raft::linalg::map(
handle,
min_cad_view,
[] __device__(raft::KeyValuePair<IndexT, DataT> pair, DataT weight) {
pair.value *= weight;
return pair;
},
raft::make_const_mdspan(min_cad_view),
batch_sw.value());
}
computeClusterCost(
handle, min_cad_view, ws, batch_inertia.view(), raft::value_op{}, raft::add_op{});
raft::linalg::add(clustering_cost.data_handle(),
clustering_cost.data_handle(),
batch_inertia.data_handle(),
1,
stream);

const bool needs_future_batch =
batch_pos + 2 < data_batches.num_batches() || seed_iter + 1 < n_init;
if (needs_future_batch) {
const auto next_batch_pos = (batch_pos + 2) % data_batches.num_batches();
data_batches.recycle(data_batch, next_batch_pos);
if (weight_batch.has_value()) { weight_batches->recycle(*weight_batch, next_batch_pos); }
} else {
data_batches.release(data_batch);
if (weight_batch.has_value()) { weight_batches->release(*weight_batch); }
}
}
raft::copy(handle,
raft::make_pinned_scalar_view(h_inertia.data_handle()),
raft::make_device_scalar_view<const DataT>(clustering_cost.data_handle()));
raft::resource::sync_stream(handle);
iter_inertia = *h_inertia.data_handle();
}

if (iter_inertia < inertia[0]) {
Expand Down
Loading
Loading