Enable kernel & memcpy overlapping in IVF index building (#230)

Currently, in IVF index building (both IVF-Flat and IVF-PQ), large dataset is usually in pageable host memory or mmap-ed file. In both case, after the cluster centers are trained, the entire dataset needs to be copied twice to the GPU -- one for assigning vectors to clusters, the other for copying vectors to the corresponding clusters. Both copies are done using `batch_load_iterator` in a chunk-by-chunk fashion. Since the source buffer is in pageable memory, the current `batch_load_iterator` implementation doesn't support kernel and memcopy overlapping. This PR adds support on prefetching with `cudaMemcpyAsync` on pageable memory. We achieve kernel copy overlapping by launching kernel first following by the prefetching of the next chunk. 

We benchmarked the change on L40S. The results show 3%-21% speedup on index building, without impacting the search recall (about 1-2%, similar to run-to-run variance). 
algo | dataset | model | with prefetching (s) | without prefetching (s) | speedup
-- | -- | -- | -- | -- | --
IVF-PQ | deep-100M | d64b5n50K | 97.3547 | 100.36 | 1.03
IVF-PQ | wiki-all-10M | d64-nlist16K | 14.9763 | 18.1602 | 1.21
IVF-Flat | deep-100M | nlist50K | 78.8188 | 81.4461 | 1.03

This PR is related to the issue submitted to RAFT: rapidsai/raft#2106

Authors:
  - Rui Lan (https://github.com/abc99lr)
  - Corey J. Nolet (https://github.com/cjnolet)

Approvers:
  - Tamas Bela Feher (https://github.com/tfeher)

URL: #230

Author: abc99lr

cpp/bench/ann/src/cuvs/cuvs_ivf_flat_wrapper.h

@@ -26,6 +26,7 @@
 #include <raft/core/resource/cuda_stream.hpp>
 #include <raft/linalg/unary_op.cuh>
 #include <raft/util/cudart_utils.hpp>
+#include <rmm/cuda_stream_pool.hpp>
 
 #include <cassert>
 #include <fstream>
@@ -96,6 +97,9 @@ class cuvs_ivf_flat : public algo<T>, public algo_gpu {
 template <typename T, typename IdxT>
 void cuvs_ivf_flat<T, IdxT>::build(const T* dataset, size_t nrow)
 {
+  // Create a CUDA stream pool with 1 stream (besides main stream) for kernel/copy overlapping.
+  size_t n_streams = 1;
+  raft::resource::set_cuda_stream_pool(handle_, std::make_shared<rmm::cuda_stream_pool>(n_streams));
   index_ = std::make_shared<cuvs::neighbors::ivf_flat::index<T, IdxT>>(
     std::move(cuvs::neighbors::ivf_flat::build(
       handle_,

cpp/bench/ann/src/cuvs/cuvs_ivf_pq_wrapper.h

@@ -30,6 +30,7 @@
 #include <raft/linalg/unary_op.cuh>
 #include <raft/neighbors/refine.cuh>
 #include <raft/util/cudart_utils.hpp>
+#include <rmm/cuda_stream_pool.hpp>
 
 #include <type_traits>
 
@@ -115,6 +116,9 @@ void cuvs_ivf_pq<T, IdxT>::load(const std::string& file)
 template <typename T, typename IdxT>
 void cuvs_ivf_pq<T, IdxT>::build(const T* dataset, size_t nrow)
 {
+  // Create a CUDA stream pool with 1 stream (besides main stream) for kernel/copy overlapping.
+  size_t n_streams = 1;
+  raft::resource::set_cuda_stream_pool(handle_, std::make_shared<rmm::cuda_stream_pool>(n_streams));
   auto dataset_v = raft::make_device_matrix_view<const T, IdxT>(dataset, IdxT(nrow), dim_);
   std::make_shared<cuvs::neighbors::ivf_pq::index<IdxT>>(
     std::move(cuvs::neighbors::ivf_pq::build(handle_, index_params_, dataset_v)))

cpp/include/cuvs/neighbors/ivf_flat.hpp

@@ -445,11 +445,18 @@ void build(raft::resources const& handle,
 /**
  * @brief Build the index from the dataset for efficient search.
  *
+ * Note, if index_params.add_data_on_build is set to true, the user can set a
+ * stream pool in the input raft::resource with at least one stream to enable kernel and copy
+ * overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
  *   // use default index parameters
  *   ivf_flat::index_params index_params;
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping. This is only applicable if index_params.add_data_on_build is set to true
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // create and fill the index from a [N, D] dataset
  *   auto index = ivf_flat::build(handle, dataset, index_params);
  * @endcode
@@ -468,11 +475,18 @@ auto build(raft::resources const& handle,
 /**
  * @brief Build the index from the dataset for efficient search.
  *
+ * Note, if index_params.add_data_on_build is set to true, the user can set a
+ * stream pool in the input raft::resource with at least one stream to enable kernel and copy
+ * overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
  *   // use default index parameters
  *   ivf_flat::index_params index_params;
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping. This is only applicable if index_params.add_data_on_build is set to true
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // create and fill the index from a [N, D] dataset
  *   ivf_flat::index<decltype(dataset::value_type), decltype(dataset::index_type)> index;
  *   ivf_flat::build(handle, dataset, index_params, index);
@@ -492,11 +506,18 @@ void build(raft::resources const& handle,
 /**
  * @brief Build the index from the dataset for efficient search.
  *
+ * Note, if index_params.add_data_on_build is set to true, the user can set a
+ * stream pool in the input raft::resource with at least one stream to enable kernel and copy
+ * overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
  *   // use default index parameters
  *   ivf_flat::index_params index_params;
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping. This is only applicable if index_params.add_data_on_build is set to true
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // create and fill the index from a [N, D] dataset
  *   auto index = ivf_flat::build(handle, dataset, index_params);
  * @endcode
@@ -515,11 +536,18 @@ auto build(raft::resources const& handle,
 /**
  * @brief Build the index from the dataset for efficient search.
  *
+ * Note, if index_params.add_data_on_build is set to true, the user can set a
+ * stream pool in the input raft::resource with at least one stream to enable kernel and copy
+ * overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
  *   // use default index parameters
  *   ivf_flat::index_params index_params;
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping. This is only applicable if index_params.add_data_on_build is set to true
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // create and fill the index from a [N, D] dataset
  *   ivf_flat::index<decltype(dataset::value_type), decltype(dataset::index_type)> index;
  *   ivf_flat::build(handle, dataset, index_params, index);
@@ -539,11 +567,18 @@ void build(raft::resources const& handle,
 /**
  * @brief Build the index from the dataset for efficient search.
  *
+ * Note, if index_params.add_data_on_build is set to true, the user can set a
+ * stream pool in the input raft::resource with at least one stream to enable kernel and copy
+ * overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
  *   // use default index parameters
  *   ivf_flat::index_params index_params;
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping. This is only applicable if index_params.add_data_on_build is set to true
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // create and fill the index from a [N, D] dataset
  *   auto index = ivf_flat::build(handle, dataset, index_params);
  * @endcode
@@ -562,11 +597,18 @@ auto build(raft::resources const& handle,
 /**
  * @brief Build the index from the dataset for efficient search.
  *
+ * Note, if index_params.add_data_on_build is set to true, the user can set a
+ * stream pool in the input raft::resource with at least one stream to enable kernel and copy
+ * overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
  *   // use default index parameters
  *   ivf_flat::index_params index_params;
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping. This is only applicable if index_params.add_data_on_build is set to true
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // create and fill the index from a [N, D] dataset
  *   ivf_flat::index<decltype(dataset::value_type), decltype(dataset::index_type)> index;
  *   ivf_flat::build(handle, dataset, index_params, index);
@@ -710,6 +752,7 @@ auto extend(raft::resources const& handle,
  * @param[in] handle
  * @param[in] new_vectors raft::device_matrix_view to a row-major matrix [n_rows, index.dim()]
  * @param[in] new_indices optional raft::device_vector_view to a vector of indices [n_rows].
+ *
  *    If the original index is empty (`orig_index.size() == 0`), you can pass `std::nullopt`
  *    here to imply a continuous range `[0...n_rows)`.
  * @param[inout] idx pointer to index, to be overwritten in-place
@@ -786,6 +829,9 @@ void extend(raft::resources const& handle,
 /**
  * @brief Build a new index containing the data of the original plus new extra vectors.
  *
+ * Note, the user can set a stream pool in the input raft::resource with
+ * at least one stream to enable kernel and copy overlapping.
+ *
  * Implementation note:
  *    The new data is clustered according to existing kmeans clusters, then the cluster
  *    centers are adjusted to match the newly labeled data.
@@ -798,6 +844,9 @@ void extend(raft::resources const& handle,
  *   index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training
  *   // train the index from a [N, D] dataset
  *   auto index_empty = ivf_flat::build(handle, index_params, dataset);
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // fill the index with the data
  *   std::optional<raft::host_vector_view<const IdxT, IdxT>> no_op = std::nullopt;
  *   auto index = ivf_flat::extend(handle, new_vectors, no_op, index_empty);
@@ -821,6 +870,9 @@ auto extend(raft::resources const& handle,
 /**
  * @brief Extend the index in-place with the new data.
  *
+ * Note, the user can set a stream pool in the input raft::resource with
+ * at least one stream to enable kernel and copy overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
@@ -829,6 +881,9 @@ auto extend(raft::resources const& handle,
  *   index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training
  *   // train the index from a [N, D] dataset
  *   auto index_empty = ivf_flat::build(handle, index_params, dataset);
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // fill the index with the data
  *   std::optional<raft::host_vector_view<const IdxT, IdxT>> no_op = std::nullopt;
  *   ivf_flat::extend(handle, dataset, no_opt, &index_empty);
@@ -850,6 +905,9 @@ void extend(raft::resources const& handle,
 /**
  * @brief Build a new index containing the data of the original plus new extra vectors.
  *
+ * Note, the user can set a stream pool in the input raft::resource with
+ * at least one stream to enable kernel and copy overlapping.
+ *
  * Implementation note:
  *    The new data is clustered according to existing kmeans clusters, then the cluster
  *    centers are adjusted to match the newly labeled data.
@@ -862,6 +920,9 @@ void extend(raft::resources const& handle,
  *   index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training
  *   // train the index from a [N, D] dataset
  *   auto index_empty = ivf_flat::build(handle, dataset, index_params, dataset);
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // fill the index with the data
  *   std::optional<raft::host_vector_view<const IdxT, IdxT>> no_op = std::nullopt;
  *   auto index = ivf_flat::extend(handle, new_vectors, no_op, index_empty);
@@ -885,6 +946,9 @@ auto extend(raft::resources const& handle,
 /**
  * @brief Extend the index in-place with the new data.
  *
+ * Note, the user can set a stream pool in the input raft::resource with
+ * at least one stream to enable kernel and copy overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
@@ -893,6 +957,9 @@ auto extend(raft::resources const& handle,
  *   index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training
  *   // train the index from a [N, D] dataset
  *   auto index_empty = ivf_flat::build(handle, index_params, dataset);
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // fill the index with the data
  *   std::optional<raft::host_vector_view<const IdxT, IdxT>> no_op = std::nullopt;
  *   ivf_flat::extend(handle, dataset, no_opt, &index_empty);
@@ -914,6 +981,9 @@ void extend(raft::resources const& handle,
 /**
  * @brief Build a new index containing the data of the original plus new extra vectors.
  *
+ * Note, the user can set a stream pool in the input raft::resource with
+ * at least one stream to enable kernel and copy overlapping.
+ *
  * Implementation note:
  *    The new data is clustered according to existing kmeans clusters, then the cluster
  *    centers are adjusted to match the newly labeled data.
@@ -926,6 +996,9 @@ void extend(raft::resources const& handle,
  *   index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training
  *   // train the index from a [N, D] dataset
  *   auto index_empty = ivf_flat::build(handle, dataset, index_params, dataset);
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // fill the index with the data
  *   std::optional<raft::host_vector_view<const IdxT, IdxT>> no_op = std::nullopt;
  *   auto index = ivf_flat::extend(handle, new_vectors, no_op, index_empty);
@@ -949,6 +1022,9 @@ auto extend(raft::resources const& handle,
 /**
  * @brief Extend the index in-place with the new data.
  *
+ * Note, the user can set a stream pool in the input raft::resource with
+ * at least one stream to enable kernel and copy overlapping.
+ *
  * Usage example:
  * @code{.cpp}
  *   using namespace cuvs::neighbors;
@@ -957,6 +1033,9 @@ auto extend(raft::resources const& handle,
  *   index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training
  *   // train the index from a [N, D] dataset
  *   auto index_empty = ivf_flat::build(handle, index_params, dataset);
+ *   // optional: create a stream pool with at least one stream to enable kernel and copy
+ *   // overlapping
+ *   raft::resource::set_cuda_stream_pool(handle, std::make_shared<rmm::cuda_stream_pool>(1));
  *   // fill the index with the data
  *   std::optional<raft::host_vector_view<const IdxT, IdxT>> no_op = std::nullopt;
  *   ivf_flat::extend(handle, dataset, no_opt, &index_empty);