[Outdated] Scaling workspace resources

cpp/bench/ann/src/raft/raft_ann_bench_utils.h

@@ -27,6 +27,7 @@
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/device_uvector.hpp>
 #include <rmm/mr/device/failure_callback_resource_adaptor.hpp>
+#include <rmm/mr/device/managed_memory_resource.hpp>
 #include <rmm/mr/device/pool_memory_resource.hpp>
 
 #include <memory>
@@ -70,13 +71,14 @@ inline auto rmm_oom_callback(std::size_t bytes, void*) -> bool
  */
 class shared_raft_resources {
  public:
-  using pool_mr_type = rmm::mr::pool_memory_resource<rmm::mr::device_memory_resource>;
-  using mr_type      = rmm::mr::failure_callback_resource_adaptor<pool_mr_type>;
+  using pool_mr_type  = rmm::mr::pool_memory_resource<rmm::mr::device_memory_resource>;
+  using mr_type       = rmm::mr::failure_callback_resource_adaptor<pool_mr_type>;
+  using large_mr_type = rmm::mr::managed_memory_resource;
 
   shared_raft_resources()
   try : orig_resource_{rmm::mr::get_current_device_resource()},
     pool_resource_(orig_resource_, 1024 * 1024 * 1024ull),
-    resource_(&pool_resource_, rmm_oom_callback, nullptr) {
+    resource_(&pool_resource_, rmm_oom_callback, nullptr), large_mr_() {
     rmm::mr::set_current_device_resource(&resource_);
   } catch (const std::exception& e) {
     auto cuda_status = cudaGetLastError();
@@ -99,10 +101,16 @@ class shared_raft_resources {
 
   ~shared_raft_resources() noexcept { rmm::mr::set_current_device_resource(orig_resource_); }
 
+  auto get_large_memory_resource() noexcept
+  {
+    return static_cast<rmm::mr::device_memory_resource*>(&large_mr_);
+  }
+
  private:
   rmm::mr::device_memory_resource* orig_resource_;
   pool_mr_type pool_resource_;
   mr_type resource_;
+  large_mr_type large_mr_;
 };
 
 /**
@@ -123,6 +131,12 @@ class configured_raft_resources {
   explicit configured_raft_resources(const std::shared_ptr<shared_raft_resources>& shared_res)
     : shared_res_{shared_res}, res_{rmm::cuda_stream_view(get_stream_from_global_pool())}
   {
+    // set the large workspace resource to the raft handle, but without the deleter
+    // (this resource is managed by the shared_res).
+    raft::resource::set_large_workspace_resource(
+      res_,
+      std::shared_ptr<rmm::mr::device_memory_resource>(shared_res_->get_large_memory_resource(),
+                                                       raft::void_op{}));
   }
 
   /** Default constructor creates all resources anew. */

cpp/include/raft/core/resource/device_memory_resource.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023, NVIDIA CORPORATION.
+ * Copyright (c) 2022-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -35,6 +35,16 @@ namespace raft::resource {
  * @{
  */
 
+class memory_resource : public resource {
+ public:
+  explicit memory_resource(std::shared_ptr<rmm::mr::device_memory_resource> mr) : mr_(mr) {}
+  ~memory_resource() override = default;
+  auto get_resource() -> void* override { return mr_.get(); }
+
+ private:
+  std::shared_ptr<rmm::mr::device_memory_resource> mr_;
+};
+
 class limiting_memory_resource : public resource {
  public:
   limiting_memory_resource(std::shared_ptr<rmm::mr::device_memory_resource> mr,
@@ -66,6 +76,29 @@ class limiting_memory_resource : public resource {
   }
 };
 
+/**
+ * Factory that knows how to construct a specific raft::resource to populate
+ * the resources instance.
+ */
+class large_workspace_resource_factory : public resource_factory {
+ public:
+  explicit large_workspace_resource_factory(
+    std::shared_ptr<rmm::mr::device_memory_resource> mr = {nullptr})
+    : mr_{mr ? mr
+             : std::shared_ptr<rmm::mr::device_memory_resource>{
+                 rmm::mr::get_current_device_resource(), void_op{}}}
+  {
+  }
+  auto get_resource_type() -> resource_type override
+  {
+    return resource_type::LARGE_MEMORY_RESOURCE;
+  }
+  auto make_resource() -> resource* override { return new memory_resource(mr_); }
+
+ private:
+  std::shared_ptr<rmm::mr::device_memory_resource> mr_;
+};
+
 /**
  * Factory that knows how to construct a specific raft::resource to populate
  * the resources instance.
@@ -144,7 +177,7 @@ class workspace_resource_factory : public resource_factory {
     // Note, the workspace does not claim all this memory from the start, so it's still usable by
     // the main resource as well.
     // This limit is merely an order for algorithm internals to plan the batching accordingly.
-    return total_size / 2;
+    return total_size / 4;
   }
 };
 
@@ -241,6 +274,21 @@ inline void set_workspace_to_global_resource(
     workspace_resource_factory::default_plain_resource(), allocation_limit, std::nullopt));
 };
 
+inline auto get_large_workspace_resource(resources const& res) -> rmm::mr::device_memory_resource*
+{
+  if (!res.has_resource_factory(resource_type::LARGE_MEMORY_RESOURCE)) {
+    res.add_resource_factory(std::make_shared<large_workspace_resource_factory>());
+  }
+  return res.get_resource<rmm::mr::device_memory_resource>(resource_type::LARGE_MEMORY_RESOURCE);
+};
+
+inline void set_large_workspace_resource(resources const& res,
+                                         std::shared_ptr<rmm::mr::device_memory_resource> mr = {
+                                           nullptr})
+{
+  res.add_resource_factory(std::make_shared<large_workspace_resource_factory>(mr));
+};
+
 /** @} */
 
 }  // namespace raft::resource

cpp/include/raft/core/resource/resource_types.hpp

@@ -42,9 +42,10 @@ enum resource_type {
   STREAM_VIEW,             // view of a cuda stream or a placeholder in
                            // CUDA-free builds
   THRUST_POLICY,           // thrust execution policy
-  WORKSPACE_RESOURCE,      // rmm device memory resource
+  WORKSPACE_RESOURCE,      // rmm device memory resource for small temporary allocations
   CUBLASLT_HANDLE,         // cublasLt handle
   CUSTOM,                  // runtime-shared default-constructible resource
+  LARGE_MEMORY_RESOURCE,   // rmm device memory resource for somewhat large temporary allocations
 
   LAST_KEY  // reserved for the last key
 };

cpp/include/raft/neighbors/detail/cagra/graph_core.cuh

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023, NVIDIA CORPORATION.
+ * Copyright (c) 2023-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -244,15 +244,19 @@ void sort_knn_graph(raft::resources const& res,
   const uint32_t input_graph_degree = knn_graph.extent(1);
   IdxT* const input_graph_ptr       = knn_graph.data_handle();
 
-  auto d_input_graph = raft::make_device_matrix<IdxT, int64_t>(res, graph_size, input_graph_degree);
+  auto large_tmp_mr = resource::get_large_workspace_resource(res);
+
+  auto d_input_graph = raft::make_device_mdarray<IdxT>(
+    res, large_tmp_mr, raft::make_extents<int64_t>(graph_size, input_graph_degree));
 
   //
   // Sorting kNN graph
   //
   const double time_sort_start = cur_time();
   RAFT_LOG_DEBUG("# Sorting kNN Graph on GPUs ");
 
-  auto d_dataset = raft::make_device_matrix<DataT, int64_t>(res, dataset_size, dataset_dim);
+  auto d_dataset = raft::make_device_mdarray<DataT>(
+    res, large_tmp_mr, raft::make_extents<int64_t>(dataset_size, dataset_dim));
   raft::copy(d_dataset.data_handle(),
              dataset_ptr,
              dataset_size * dataset_dim,
@@ -323,6 +327,7 @@ void optimize(raft::resources const& res,
 {
   RAFT_LOG_DEBUG(
     "# Pruning kNN graph (size=%lu, degree=%lu)\n", knn_graph.extent(0), knn_graph.extent(1));
+  auto large_tmp_mr = resource::get_large_workspace_resource(res);
 
   RAFT_EXPECTS(knn_graph.extent(0) == new_graph.extent(0),
                "Each input array is expected to have the same number of rows");
@@ -338,15 +343,16 @@ void optimize(raft::resources const& res,
     //
     // Prune kNN graph
     //
-    auto d_detour_count =
-      raft::make_device_matrix<uint8_t, int64_t>(res, graph_size, input_graph_degree);
+    auto d_detour_count = raft::make_device_mdarray<uint8_t>(
+      res, large_tmp_mr, raft::make_extents<int64_t>(graph_size, input_graph_degree));
 
     RAFT_CUDA_TRY(cudaMemsetAsync(d_detour_count.data_handle(),
                                   0xff,
                                   graph_size * input_graph_degree * sizeof(uint8_t),
                                   resource::get_cuda_stream(res)));
 
-    auto d_num_no_detour_edges = raft::make_device_vector<uint32_t, int64_t>(res, graph_size);
+    auto d_num_no_detour_edges = raft::make_device_mdarray<uint32_t>(
+      res, large_tmp_mr, raft::make_extents<int64_t>(graph_size));
     RAFT_CUDA_TRY(cudaMemsetAsync(d_num_no_detour_edges.data_handle(),
                                   0x00,
                                   graph_size * sizeof(uint32_t),
@@ -466,14 +472,16 @@ void optimize(raft::resources const& res,
                                   graph_size * output_graph_degree * sizeof(IdxT),
                                   resource::get_cuda_stream(res)));
 
-    auto d_rev_graph_count = raft::make_device_vector<uint32_t, int64_t>(res, graph_size);
+    auto d_rev_graph_count = raft::make_device_mdarray<uint32_t>(
+      res, large_tmp_mr, raft::make_extents<int64_t>(graph_size));
     RAFT_CUDA_TRY(cudaMemsetAsync(d_rev_graph_count.data_handle(),
                                   0x00,
                                   graph_size * sizeof(uint32_t),
                                   resource::get_cuda_stream(res)));
 
-    auto dest_nodes   = raft::make_host_vector<IdxT, int64_t>(graph_size);
-    auto d_dest_nodes = raft::make_device_vector<IdxT, int64_t>(res, graph_size);
+    auto dest_nodes = raft::make_host_vector<IdxT, int64_t>(graph_size);
+    auto d_dest_nodes =
+      raft::make_device_mdarray<IdxT>(res, large_tmp_mr, raft::make_extents<int64_t>(graph_size));
 
     for (uint64_t k = 0; k < output_graph_degree; k++) {
 #pragma omp parallel for

cpp/include/raft/neighbors/detail/cagra/utils.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023, NVIDIA CORPORATION.
+ * Copyright (c) 2023-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -176,8 +176,11 @@ class device_matrix_view_from_host {
     device_ptr = reinterpret_cast<T*>(attr.devicePointer);
     if (device_ptr == NULL) {
       // allocate memory and copy over
-      device_mem_.emplace(
-        raft::make_device_matrix<T, IdxT>(res, host_view.extent(0), host_view.extent(1)));
+      // NB: We use the temporary "large" workspace resource here; this structure is supposed to
+      // live on stack and not returned to a user.
+      // The user may opt to set this resource to managed memory to allow large allocations.
+      device_mem_.emplace(make_device_mdarray<T, IdxT>(
+        res, resource::get_large_workspace_resource(res), host_view.extents()));
       raft::copy(device_mem_->data_handle(),
                  host_view.data_handle(),
                  host_view.extent(0) * host_view.extent(1),